On October 21, 2026, GitHub detailed on its engineering blog how it is moving Git repositories to Azure and preparing its storage for the growth in agent-generated traffic. The post is written by Eleanor White and Victoria Dye.
The problem starts with the Spokes quorum, which waits for all healthy replicas to answer and locks the whole repository on every write. With a replica in Azure, the network round trip is around 25 ms, and each write needs several trips. In GitHub's words, "the slowest healthy replica sets the latency of every write".
The fix has two parts. The first is a per-reference quorum: serialization now happens per branch or tag, since writes to different references are usually independent. The second is an ephemeral virtual write replica that bridges to Azure Blob Storage and only acknowledges a write once it is durable in object storage, while also batching transactions on disjoint references. There are also virtual read replicas, created on demand and outside the quorum, which decouples read and write scaling.
GitHub cites numbers: 99.99999% of repositories already have a copy in Azure, Spokes handles more than 100 billion operations per day, most repositories see 100 to 1,000 reads for every write, and write throughput across different references was 3 to 4 times higher in internal tests.
Original source: GitHub Blog