GitHub System Design: Rebuilding Git for AI Agents

Pushes to GitHub went from 690 million a month to 3.35 billion in one year, and most of that is AI agents committing after every step. Here's the GitHub system design behind git push, why it hit a wall, and how they're rebuilding it. You'll see what a push actually does (send objects, then a compare-and-swap on one ref) and why writes are the hard part. Then how Spokes stores every repo as five full copies with a three-phase commit and a quorum, and why that couples durability to read capacity. Then the two fixes: only the ref update needs consensus, and storage moves to Azure Blob Storage while compute workers cache on demand and separate workers handle GC. GitHub reports up to 35x write throughput in internal benchmarks. I'll also cover what that number doesn't tell you yet. Chapters 0:00 Agents are flooding git push 1:09 What a push really is 2:11 How Spokes works today 2:45 The coupling problem 3:50 Minimize coordination 4:40 Split storage from compute 6:06 Tradeoffs and what's next 7:00 What to steal for your systems Links GitHub's post: https://github.blog/engineering/archi... git update-ref docs: https://git-scm.com/docs/git-update-ref git repack docs: https://git-scm.com/docs/git-repack git gc docs: https://git-scm.com/docs/git-gc Azure Blob Storage docs: https://learn.microsoft.com/azure/sto... #systemdesign #git #github