GitHub Pushes 99.99999% of Repositories to Azure Over Agent Traffic
Surging traffic from autonomous coding agents pushed GitHub to duplicate almost all repositories across on-prem hardware and Microsoft Azure, though latency still locks out its busiest codebases.
GitHub announced it has mirrored 99.99999% of its repositories into Microsoft Azure, keeping copies in both its legacy data centers and the public cloud. The platform attributed the shift to what it calls an "agentic shift," where automated coding bots and agents have triggered an exponential surge in Git traffic.
According to the engineering team, the migration began in August 2025 when the first copy landed in Azure's Central US region. Most repositories now rely on five replicas, with two hosted in Azure. That cloud setup already handles roughly half of GitHub's total Git traffic.
The remaining fraction of a percent represents some of the platform's most heavily hit projects. Moving them stalled on a hard physical limit: a roughly 25-millisecond network round-trip between GitHub's private facilities and the target Azure region 1,600 kilometers away. Because its replication coordinator, Spokes, waits on quorum acknowledgments across replicas, multi-round-trip write operations quickly degrade throughput on hyperactive repositories.
To work around the bottleneck, GitHub is overhauling how Spokes handles writes. Instead of locking an entire repository, the system serializes updates at the individual reference level. In internal testing, writes targeting independent refs saw throughput increase three- to fourfold with lower latency.
The company is also decoupling compute for reads and writes, introducing virtual write replicas backed directly by Azure Blob Storage. While Spokes currently processes more than 100 billion operations daily, engineers noted that raw write throughput remains the ultimate constraint: every write still updates heavy full-repository replicas, even when only a fraction of that data demands strict global ordering.
Source: habr via GitHub Engineering
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.