OpenAI Releases Pando Git Backend for Agent-Scale Parallel Pushes

OpenAI released Pando, a Git-compatible backend built to handle agent-scale parallel pushes, after Codex pushed its monorepo to concurrent multi-branch writes and massive parallel reads. Pando provides branch-level causal consistency and reuses packfiles across clients, cutting median push latency in production to about one-fifth of the previous level and serving nearly 1 PB of Git data per day.
Event version 1.
Key insights
Codex overloaded the old Git system
Codex amplified parallel changes, branches, pushes and CI pulls simultaneously. Since late last year the old system suffered slow pushes and pulls and merge failures, with reliability at one point below 90%.
Pando relaxes two consistency assumptions
Pando makes a single ref serializable and keeps only causal consistency across the repository, rather than requiring whole-repo consistency at a moment in time. Packs go into object storage, with metadata recording tips, dependencies and pack pointers.
Writes advertise only the main branch as base
By default writes advertise only the trunk as base, so advertisements do not grow with millions of branches and packs are easier to reuse, at the cost of extra transfer when stacking branches. Median pack size in the repository is under 20 KiB and P99 under 3 MiB.
Reads reuse stored packs
The read side maps have/want to stored ref versions and delivers cacheable packs via packfile-uri, with Express hit rate above 99% and standard machines saturating 10 Gbps. Background compaction follows an LSM approach, merging L0 up to L4 as both compression and a skip index.
Ten weeks from project start to production
Pando was rewritten in Rust and reached production in about 10 weeks using Codex, starting to take traffic around April, first through a dual-write proxy and then phased traffic shifts. It sent more than 57 TB in a single hour with over 99% local cache hits, and improvements were fed back upstream.