GitHub changes its Git write quorum to move busy repositories into Azure

GitHub says that most repositories now have two of their five replicas in Azure, and Azure serves about half of Git traffic. The remaining busiest repositories were held back by cross-region round trips of around 25ms. The quorum used to wait for every healthy replica before moving on, so the slowest replica set the latency and throughput ceiling for each write.

The new approach limits serialization to writes that touch the same reference, so pushes to different branches no longer wait on each other. Separately, GitHub is decoupling storage from compute. A virtual write replica joins the quorum, streams writes to Azure Blob Storage, and serves no reads, while virtual read replicas warm their caches from object storage without joining the quorum.

Reads outnumber writes heavily, often 100 to 1,000 to one, so GitHub expects read scaling to benefit from this split. The quorum still limits write throughput, and GitHub says a follow-up post covering the end state is due next week, around GitHub Universe.

Agents and developers pushing to their own branches in the same repository should see less queueing, though each write still pays cross-region latency.

Writes that update the same reference still wait for each other, so agents that repeatedly target one shared branch may still contend for throughput.

Read-heavy agent workloads may get more capacity from virtual read replicas, but reads served from them reconcile with object storage in the background, so they may not suit every read pattern.