GitHub details 'Rubicon', a Git storage redesign aimed at thousands of agent pushes per second
In a company engineering post, GitHub says its new Git storage architecture separates reference coordination from object storage and handled 7,200 pushes per second in internal tests. It is in limited private preview, with a broader rollout planned early next year.
GitHub has published a technical account of its new Git storage architecture, which it announced at its Universe conference and has nicknamed "Rubicon." The company frames the work as preparation for AI agents that push code to the same repository in parallel and around the clock. [1]
The headline figures come from GitHub's own internal testing and have not been independently verified. GitHub says the system supported 5,000 concurrent agents on one repository, sending 7,200 pushes per second, while read latency stayed consistently low. [1]
GitHub says the busiest repositories on GitHub.com currently see at most dozens of writes in a single second, usually in short bursts. It expects those bursts to become a sustained rate within the next year, and it says it set a design target of at least two orders of magnitude more scale. [1]
The design rests on a simple observation: Git objects are immutable and named by their contents, so only reference updates need agreement between writers. Under the new design, packs and read-acceleration data live in Azure Blob Storage. References and the pack manifest are held in a separate replicated log service that GitHub calls Ref-CAS, short for Reference Compare-and-Swap. [1]
Requests still enter through GitHub's existing Spokes routing layer. From there they go to "virtual replicas," disposable compute whose local disk serves only as a rebuildable cache. GitHub says it adds virtual replicas when a repository gets busy. [1]
According to the post, GitHub first tried keeping references in Blob Storage. A single-blob log reached write throughput only in the mid-hundreds per second. Splitting references across multiple blobs raised write capacity but slowed reads. Ref-CAS, by contrast, commits each reference update as a quorum compare-and-swap that locks only the references it touches. GitHub says it can handle reference state alone at tens of thousands of operations per second. [1]
Maintenance has also moved off the serving path. In GitHub's legacy design, compaction ran on every replica that served traffic. In the new design, separate compaction nodes merge packs in the background, each claiming shards through leases. GitHub says this has kept latency flat as writes climb. [1]
GitHub says developers will push, fetch and branch exactly as before. The architecture is in a limited private preview, with a broader rollout planned for early next year. Planned follow-ups include global replicas, partial replicas, an atomic multi-file commit API and event streams for agents. [1]
Sources
- GitHub: Inside GitHub’s new Git storage architectureSource published · TechDots found it
Statements labelled as a company's claim have not been independently confirmed by TechDots.
Get Technology stories by email
Related coverage
Spotify Engineering questions the idea of a single 'AI product manager' role
A Nov. 9 post on Spotify's engineering blog examines AI product management from three angles, starting from the observation that most coverage treats 'the AI PM' as one job. Only an excerpt was available for review.
Stripe says it moved $1B in vendor payments onto its own Global Payouts product
In a company blog post, Stripe says it began shifting its vendor payments to Global Payouts in early 2026, cutting new-market setup from six weeks to one with a 99.9% payout success rate. The figures are Stripe's own and have not been independently verified.
Stripe says users can now query their Stripe data in natural language through an agent
In a 5 November blog post, Stripe said customers can use natural language with an agent to analyse their Stripe data. Its examples include building an investor slide deck and a custom revenue dashboard. Details beyond the announcement are not available.