Scaling Git at OpenAI
Points and comments are a snapshot, not live.
OpenAI built a Git backend, Pando, to scale writes and reads across its monorepo.
OpenAI's monorepo growth, accelerated by Codex-driven parallel development, exceeded conventional Git backends. They built Pando, a new backend that uses causal consistency and packfile reuse to scale writes and reads without changing the Git client. Pando coordinates branch updates independently for parallel pushes and serves cached packfiles for high read throughput. Median push latency dropped nearly 5x. Pando delivers nearly a petabyte of Git data per day across OpenAI's monorepo.
What commenters are saying
Several commenters questioned the necessity and sustainability of such high code volume, with some attributing it to AI agents producing 'slop' and bloated commits and tests. Others argued the volume is natural for an AI lab where agents write most code and humans don't need to review every line. Some praised the blog post as a technical deep-dive free from hype, while others debated whether the code explosion is an asset or liability for long-term maintainability. Skeptics noted agent-generated code often requires repeated fix commits.