GitHub has detailed a new Git storage architecture, called Project Rubicon, that separates immutable repository objects from the small reference updates that must be ordered. The redesign targets a workload that traditional repository servers were not built for: thousands of coding agents pushing to the same repository at once while CI systems and other agents multiply read traffic.
What changed
Git objects and pack indexes live durably in Azure Blob Storage. Branch and tag references, plus the manifest of packs that form a repository, live in a replicated compare-and-swap log called Ref-CAS. Disposable virtual replicas handle reads and writes with local disks treated as rebuildable caches, while separate workers compact packs and perform garbage collection.
A push uploads its pack before committing a small Ref-CAS transaction that updates only the references it touches. Independent branches can therefore advance in parallel. Reads first compare their cached reference version with Ref-CAS, then fetch only the packs needed to catch up.
The practical consequence
Git clients keep the same push, fetch, branch, and protection behavior, but GitHub can add serving replicas, compaction workers, and storage capacity independently. That reduces the chance that a very active agent-driven repository becomes limited by one repository server or by maintenance work competing with live traffic.
The design also changes failure recovery: correctness no longer depends on a request reaching a particular replica, and a replacement can rebuild its cache from durable object storage and reference state.
Limits and evidence
Concurrent updates to the same reference still contend, so only one matching compare-and-swap can win. Cold caches may need to retrieve packs from object storage, and high write rates require background compaction to prevent pack counts from growing without bound. GitHub describes global replicas, partial replicas, per-object reads, materialized views, and scale-to-zero behavior as future work, not current capabilities.
GitHub reports internal tests with 5,000 concurrent agents and 7,200 pushes per second to one repository, plus reference transactions in the tens of thousands per second. These are vendor-run internal results, not benchmarks independently reproduced by DrComps, and GitHub has not published the full workload, hardware, cost, or failure-testing methodology. The company says it is building the architecture underneath GitHub, but does not state that every repository has completed migration.