GitHub has published the design of Rubicon, the Git storage architecture it announced at GitHub Universe, after internal tests carried 5,000 concurrent agents on one repository at 7,200 pushes per second while serving reads with steady latency. The busiest repositories on GitHub today take at most dozens of writes per second, usually in short bursts.

The gain comes from separating stored code from branch pointers. A push uploads immutable packs to Azure Blob Storage, and a replicated log service called Ref-CAS holds the references and the pack manifest. Each reference update is a compare-and-swap that locks only the references it touches, so agents pushing to their own branches commit in parallel and only simultaneous updates to the same reference serialize. A replica that is already current answers a read with a single version comparison, and Ref-CAS transacts tens of thousands of operations per second because it stores pointers, never object data. Compaction and garbage collection run on separate nodes against durable storage instead of competing with the servers that serve traffic.

Rubicon is in a limited private preview, with a broader rollout planned for early next year.

Related: GitHub Rewrites Copilot Runtime in Rust Using Copilot Itself

GitHub Blog · GitHub on X