GitHub Enhances Git Storage for Agent-Driven Development

What happened

GitHub introduced a new Git storage architecture at GitHub Universe, designed to manage thousands of writes per second and scale reads as needed. The system supports 5,000 concurrent agents working on a single repository during internal testing. It also enables 7,200 pushes per second to a repository, maintaining low latency for read operations.

The new architecture is based on two core principles: minimizing coordination and decoupling storage from compute. GitHub moved 99.99999% of repositories into Azure as part of this redesign. The system uses virtual replicas backed by Azure Blob Storage, allowing for independent scaling of different repository components.

Why it matters

The shift is driven by the growing impact of agent-driven development on GitHub, which has significantly altered the behavior of busy repositories. By decoupling storage from compute, the architecture ensures that developers can continue pushing, fetching, and branching without changes to their workflows. This enhances reliability and scalability across the platform.

Key details

The architecture scopes write quorums to individual references, enabling different parts of the Git infrastructure to scale independently. It also allows virtual replicas to catch up by downloading only necessary data, minimizing commit latency. The system ensures that reads remain efficient when there are no changes, which is common due to the read-to-write ratio.

GitHub introduced the new design, called 'Project Rubicon,' in May 2026. The system uses Spokes to handle routing, load balancing, and internal API requests for repository data. Compaction nodes work in the background against Blob Storage and Ref-CAS to merge packs and remove unreachable objects without affecting virtual replicas.