On 2026-11-05 GitHub Engineering detailed the next-gen Git storage announced at Universe (codename Rubicon). Vendor internal tests: ~5,000 concurrent agents and ~7,200 pushes/sec on one repo with low-latency reads. Objects live in Azure Blob; reference updates go through Ref-CAS; disposable virtual replicas serve compute. Limited private preview; broader rollout early next year.

Key Takeaways

  • ✓Source: GitHub Blog Inside GitHub's new Git storage architecture (2026-11-05) + @github; prior series Building… and The bridge…
  • ✓Scale (vendor internal tests): ~5,000 concurrent agents and ~7,200 pushes/sec on one repo with low-latency reads
  • ✓Architecture: objects/packs in Azure Blob; refs + pack manifest in Ref-CAS; virtual replicas scale out; compaction/GC on separate nodes
  • ✓Principles: coordinate only reference updates; decouple storage from compute so adding read capacity does not slow writes
  • ✓Availability: limited private preview; broader rollout early 2027; Git workflows unchanged for developers
GitHub Project Rubicon: New Git Storage Hits ~5,000 Concurrent Agents and ~7,200 Pushes/sec in Internal Tests
🖼️Official Media
Click to view high-res
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points

Agents push busy repos from short write bursts into sustained concurrency: each agent waits on a push, and every write fans out to CI and other agents. GitHub says platform Git activity more than doubled from Sep 2025 to Aug 2026, with ~7.38B commits in Sep 2026. Legacy design couples durable replicas with read scale, so adding readers slows writers.

Architecture Highlights & Internals

Project Rubicon uploads immutable objects to Azure Blob first, then Ref-CAS commits a quorum CAS only on touched refs—most agents push private branches. Spokes routes onto disposable virtual replicas with rebuildable disk caches. Compaction and GC run on separate nodes against Blob/Ref-CAS.

Authoritative Benchmarks & Measured Scores

Vendor internal tests only: ~5,000 concurrent agents and ~7,200 pushes/sec on one repo with low-latency reads; an earlier series post cites up to ~35× write throughput. No third-party replication. Early Blob conditional-write prototypes topped mid-hundreds of writes/sec; Ref-CAS is described as tens of thousands of ref transactions/sec.

Developer Hands-on Guide

Push, fetch, branch, and protections stay the same. Limited private preview now; broader rollout aimed for early 2027. Follow-ons include clone-free file reads, an atomic multi-file commit API, event streams, and lighter clones. Follow the three-post Blog series; do not treat internal stress numbers as your repo SLA.

Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

Up NextGitHub Octoverse 2026: agents start over 40% of pull requests, pushes hit 17.9 billion, and Copilot users tapped 47 models
Read Now→