R/place for Git – Open-source Git server sustaining 567 pushes/s

Aspen is an open-source Git server from Reason Machines, a Y Combinator startup, built for repositories that many coding agents push to at once. The founder's Hacker News post says that in the company's experiments, giving 100 agents push access was enough to overwhelm ordinary Git servers. The launch blog explains why: a classic server takes every push to a repository through one queue, so a push to one branch waits behind pushes to every other branch, and it builds a fresh pack, Git's compressed bundle of files and commits, for every clone and fetch. Aspen follows the design OpenAI described for its internal Pando backend in Scaling Git at OpenAI. Each branch is a row in Postgres that moves with a compare-and-swap, meaning the server moves the branch from commit A to commit B only if it still points at A. Pushes to different branches never wait for each other, and when two pushes hit the same branch one wins and the other is told to fetch first, as stock Git already does. Fetches are answered from the packs exactly as they were pushed, kept in object storage, instead of being recompressed for each client. Clients use ordinary git over HTTPS or SSH.

That design is why Aspen's headline number is not a like-for-like win over Cursor's Git server, Origin. Reason Machines reports 567 pushes a second sustained for five minutes with no errors, on one 8-core server with S3 Standard storage and a 4-core Postgres database, pushing to a copy of the git.git repository, with each push stored before Git reported success. Cursor's published figures are 120 pushes a second on S3 Standard and more than 300 on the faster S3 Express One Zone, from a replicated cluster; Reason Machines notes they come from Cursor's own monorepo, whose size the post does not give. Cursor also runs every push to a repository through one log so that every reader sees one consistent view of all branches. Aspen gives that up: a reader looking at several branches may see each at a slightly different moment. Reason Machines adds that each push costs more as a repository grows, so very large monorepos see lower rates today, and that separate landings on main still run one at a time, at about 4 a second.

The project is pre-alpha. Its README says the speed figures come from a single run on one server and should be treated as indicative rather than as performance claims. A 24-hour chaos test with no lost acknowledged push and a teammate's reproduction of the benchmarks are still unfinished launch-gate items, and the README advises against storing data you cannot afford to lose. Aspen has no web interface or partial clones yet, and because a push sees only the trunk branch, force pushes and branch deletes need an extra -o force option.