Pando
A Git-compatible version control backend developed by OpenAI that preserves the Git protocol and object model while redesigning the storage and concurrency layers to support large-scale parallel commits by AI agents.
Timeline (last 90 days)
OpenAI rebuilt the Git server Pando for Codex, reducing median push latency by nearly 80% and handling about 1PB of data per day.
Pando's design includes making the processing of each branch independent of the others.
OpenAI developed its own Git backend, Pando, for its ultra-large monorepo.
Pando reduced median push latency by nearly 80%.
Pando carries nearly petabyte-scale Git data transfer per day.
The industry's approaches to the version-control pressure of monorepos fall roughly into two categories: virtual file systems (such as VFS for Git and some of Sapling's ideas) that materialize files on demand; and rewriting server-side transfer and storage (Pando belongs to the latter).
Pando reuses already-stored packfiles across multiple fetches.
The nearly 80% latency reduction is a median measured under OpenAI's own workload, and whether it can be reproduced when migrated to other teams depends on whether the access patterns are similar.
Pando currently mainly serves OpenAI internally.
OpenAI built Pando, a Git-compatible backend system, to cope with the large-scale push pressure generated by agent tools such as Codex.
8 more timeline events
All Facts (18)
OpenAI internally uses a giant monorepo to manage a large amount of code.
60%UnverifiedOpenAI rebuilt the Git server Pando for Codex, reducing median push latency by nearly 80% and handling about 1PB of data per day.
50%UnverifiedPando reduced median push latency by nearly 80%.
50%UnverifiedPando carries nearly petabyte-scale Git data transfer per day.
50%UnverifiedPando reuses already-stored packfiles across multiple fetches.
50%UnverifiedThe industry's approaches to the version-control pressure of monorepos fall roughly into two categories: virtual file systems (such as VFS for Git and some of Sapling's ideas) that materialize files on demand; and rewriting server-side transfer and storage (Pando belongs to the latter).
50%UnverifiedPando's design includes making the processing of each branch independent of the others.
50%UnverifiedThe nearly 80% latency reduction is a median measured under OpenAI's own workload, and whether it can be reproduced when migrated to other teams depends on whether the access patterns are similar.
50%UnverifiedOpenAI developed its own Git backend, Pando, for its ultra-large monorepo.
50%UnverifiedPando currently mainly serves OpenAI internally.
50%UnverifiedOpenAI built Pando, a Git-compatible backend system, to cope with the large-scale push pressure generated by agent tools such as Codex.
50%UnverifiedPando remains compatible at the Git protocol level, so existing toolchains and clients can continue to work without major modifications.
50%UnverifiedPando optimizes the push path, possibly involving improved reference updates, reduced lock contention or a restructured object storage write process.
50%UnverifiedThe OpenAI engineering team developed Pando, a Git-compatible backend designed specifically for the development scale that Codex can unlock.
50%UnverifiedPando's choice of Git compatibility acknowledges Git's status as the de facto standard: even with an underlying rebuild, the interface layer is not readily changed.
50%UnverifiedIn the short term, Pando is part of OpenAI's internal infrastructure and is inaccessible to ordinary developers.
50%UnverifiedOpenAI has not disclosed Pando's specific implementation details, performance data, or deployment scale.
50%UnverifiedPando is not positioned to rebuild the version control system, but to retain Git compatibility and replace the backend architecture to withstand large-scale concurrent generation.
50%