99.99999% of its repositories now have copies in both company data centers and Microsoft Azure. Most repositories keep five replicas, with two in Azure, and the cloud platform currently serves about half of Git traffic.
The remaining fraction contains some of GitHub’s busiest repositories. A roughly 25-millisecond round trip between regions slows writes because the current quorum waits for every healthy replica while holding a repository-wide lock. Busy projects can reach their throughput ceiling before a remote write finishes.
GitHub is narrowing serialization from an entire repository to each Git reference, such as an individual branch. Updates that touch different references can then proceed independently, while competing writes to the same reference still remain ordered. That pattern matters as developers and coding agents create more parallel branches.
The migration separates two risky changes: GitHub moved repository data first and is now evolving its storage architecture. The company also plans stateless read replicas backed by object storage, allowing read capacity to scale separately from compute. The last high-traffic repositories still depend on these changes being deployed successfully.