GitHub's Project Rubicon Claims 7,200 Pushes a Second to One Repo, But Only in Internal Tests and a Private Preview
Software / news
GitHub's Project Rubicon Claims 7,200 Pushes a Second to One Repo, But Only in Internal Tests and a Private Preview
The new Git storage design moves packs into Azure Blob Storage and ref updates into a quorum log called Ref-CAS. GitHub's own post says the busiest repositories today see dozens of writes a second.

GitHub has rebuilt how it stores Git repositories around Azure Blob Storage and a new reference log, and says internal tests reached 5,000 concurrent agents and 7,200 pushes a second on a single repository. The design, internally called Project Rubicon, is in limited private preview, with a broader rollout planned for early 2027, according to GitHub's engineering post dated Nov. 5, 2026.
The post is by Karthik Raman, a Distinguished Engineer, and Brian Celenza, a Principal Software Engineer. Both figures are GitHub's own and come from internal testing; no outside party has reproduced them.

What 7,200 pushes a second is measured against
The baseline is the part of the post that does the work. GitHub says its busiest repositories today see at most dozens of writes a second, usually in short bursts, and that the target is at least two orders of magnitude above that.
The pressure is agents. A chart in the post puts one repository's daily requests at 44 million on the March 2026 average and 181 million in May, about four times as many.
- March 202644 million requests
- May 2026181 million requests
Source: GitHub engineering blog, Nov. 5, 2026, accessed 2026-11-07
The Register reported the same direction on a wider scale. Citing GitHub's figures, reporter Joab Jackson wrote on Oct. 9 that GitHub events rose from 218.2 billion a month in September 2025 to 473.3 billion in August 2026, and that commits reached 7.38 billion in September 2026, five times the year before.
How the old and new designs differ
According to The Register, the old system, Spokes, uses a three-phase commit and keeps full repository copies on several local disks, so every push waits on the slowest replica needed for quorum. The Register also counted 10 performance-degrading incidents in April 2026 and nine in May, citing GitHub's Availability Report.
| Spokes (old) | Rubicon (new) | |
|---|---|---|
| Pack data | Full copies on local disks | Azure Blob Storage |
| Write path | Quorum of replicas per push | One blob write, then one Ref-CAS transaction |
| Reads | Same replicas as writes | Disposable virtual replicas with a rebuildable cache |
| Coordination | Every write | Reference updates only |
The post gives the reasoning in one line: "very little of a write needs to be ordered." Git objects are immutable and named by content, so writers can store them concurrently. Only reference updates, the pointers that say which commit a branch is on, need agreement.
That agreement lives in Ref-CAS, a replicated log where each reference update is a compare-and-swap committed by a majority quorum. GitHub says it handles tens of thousands of reference transactions a second. A virtual replica uploads a push's pack to Blob Storage first and commits to Ref-CAS second, so a reference never points at objects that are not yet durable.
The two designs GitHub threw away
The post names two dead ends. A write-ahead log kept in a single blob capped writes in the mid-hundreds a second. Splitting references across segment blobs raised write throughput but made reads as slow as the slowest segment.
Compaction runs on separate nodes holding per-shard leases. They merge packs, drop unreachable objects and verify that a merged pack contains the same objects as its inputs before publishing it.
What is not shown
The post gives no latency distribution, no pack-size assumptions for the 7,200 figure and no cost comparison with Spokes. The Register's 35 times figure for write throughput is also GitHub's, from early internal tests, and the two numbers describe different tests. The Register said Celenza gave no migration timeline in the earlier announcement.
Related coverage on this site covers other stores built on object storage: Rivet's S3-backed agent orchestrator acknowledges writes from local SSDs rather than S3, and NVIDIA's NVKE preview shows how a preview can differ from its pitch.
The post lists single-object fetches, partial replicas and an atomic multi-file commit API as planned work. The next dated milestone is the broader rollout GitHub places in early 2027.
Sources
More in Software
- 01Zigzag, a 35 MB Mac Browser, Courts Arc Users 35 Days Before ShutdownThe WebKit-based browser imports Arc spaces and pinned tabs, but its README lists no sync, no Boosts and a partial Chrome extension API.
- 02Nimony 0.6.2 Sets a Year-End Goal for Nim 3 With 44 Days LeftThe compiler that will become Nim 3 is MIT-licensed and builds natively without LLVM, but its FAQ says it cannot yet compile Nim 2 code unchanged.
- 03Alloy CSS Runs Menus and Tabs Without JavaScript, but Dialogs Need Safari 26.2The 24 KB MIT stylesheet leans on popover and invoker commands, so its no-script promise has browser floors that differ by component.
- 04Windows 11 Telemetry Flags Which Coding Agents You Run, Researcher FindsXusheng Li traced a service-experiment feature in build 26200.9974 that matches executable names such as claude.exe and queues the result as optional diagnostic data.