GitHub Rebuilds Git Underlying Architecture: 5,000 AI Agents Work in Parallel, Single Repository Processes 7,200 Code Pushes Per Second

GitHub unveiled its next-generation Git storage architecture, codenamed Rubicon. Internal tests show that 5,000 AI agents can work simultaneously in a single code repository, processing 7,200 code pushes per second while maintaining low read latency. Rubicon is currently in limited private beta, with plans to expand rollout in early 2027.

To handle massive parallel commits, GitHub separates "storing code" from "updating branches." Pushes first save newly added or modified objects, then update branch pointers. Previously, shared update records caused contention, while splitting records slowed down reads. The newly developed Ref-CAS service allows different branches to update concurrently, handling contention only when conflicts occur on the same branch, and maintains a unified version number for the entire repository, enabling other programs to quickly detect changes via this version number. Background dedicated machines merge accumulated data packets to prevent read slowdowns.

[BlockBeats]