OpenAI Cuts Git Push Latency Nearly 5x With Pando, Serving 1 Petabyte a Day
Updated
Updated · OpenAI · Nov 2
OpenAI Cuts Git Push Latency Nearly 5x With Pando, Serving 1 Petabyte a Day
1 articles · Updated · OpenAI · Nov 2
Summary
Pando now powers all development in OpenAI’s monorepo, where the new Git backend cut median push latency by nearly 80% and improved merge reliability.
Codex-driven parallel development had overwhelmed the previous system—slow fetches, failed merges and reliability at times below 90%—pushing OpenAI to rebuild the backend without changing Git workflows.
Pando scales writes by coordinating branch updates independently and scales reads by reusing stored packfiles, letting thousands of clients push concurrently while clone and fetch traffic can saturate 10 Gbps links.
Built in 10 weeks and serving production traffic since April, the system now delivers nearly 1 petabyte of Git data per day and has served more than 57 terabytes of packs in a single hour.
OpenAI said it is upstreaming related Git improvements and is exploring ways to share Pando’s technology beyond its own monorepo as agent-driven software work expands.