The bridge to agent-scale Git infrastructure

GitHub moved a copy of 99.99999% of repositories from on-premises data centers into Azure to absorb massive, agent-driven traffic growth while hitting limits on power, space, and supply-chain capacity. To avoid changing storage architecture and hosting platform simultaneously, the data was migrated first: the first Azure replica arrived in August 2025, most repositories now have two of five replicas in Azure, and Azure handles roughly half of Git traffic. Replication uses five replicas for durability and availability, and Spokes’ preferential routing directs reads to a local replica to minimize latency. A tiny fraction of the busiest repositories remained tied to the old region because a 25 ms round trip and the current quorum serialization lowered their write throughput.

The bridge between the old and new designs is a change in how quorums serialize writes - per-reference quorums - which lets those high-traffic repositories move into Azure and enables replicas that keep no persistent state. Decoupling storage from compute and relying on object storage lets reads scale independently of replica state, while minimizing coordination reduces cross-region contention. Sequencing the migration (move data first, then evolve architecture) kept risk low and allowed cloud capacity to absorb growth immediately, paving the way for further architectural evolution once data was colocated in Azure.