Inside GitHub’s new Git storage architecture
GitHub has introduced a new Git storage architecture designed to handle large-scale agent-driven development. The architecture can support thousands of writes per second, thousands of concurrent agents, and 7,200 pushes per second to a repository while maintaining low latency. Key components include virtual replicas, decoupling storage from compute, and minimizing coordination. This allows the storage and serving of Git data to scale independently, reducing the need for agreement and order in writes. The architecture is designed to handle the increased traffic and scale of agent-driven development, with a target of at least two orders of magnitude more scale than current levels. The system can support multiple writer nodes, reducing the cost and complexity of indexing and checking object connectivity. This new architecture will enable GitHub to support hundreds of millions of developers and their agents.