GitHub details per-reference write serialization to move its busiest repositories to Azure despite 25ms cross-region latency

GitHub says 99.99999% of repositories now live in both its data centers and Azure, with most repos using two of five replicas in Azure and Azure serving about half of Git traffic. The remaining busiest repositories are constrained by a 25ms data-center-to-Azure round trip, so GitHub is shifting from repository-wide write serialization to per-reference serialization, which showed 3–4x higher throughput at lower latency in internal tests. GitHub also outlines planned virtual replicas backed by Azure Blob Storage to separate read scaling from write quorum behavior.

For teams building Git-scale storage, the reported 100–1,000 reads per write and 25ms cross-region penalty show why a single repository-wide lock can turn latency into a throughput ceiling. GitHub’s approach suggests prioritizing contention scope before adding replicas: per-reference coordination can preserve ordering where needed while allowing independent branch writes to proceed, but the gains depend on workloads touching different references and are based on GitHub’s internal tests.