---
title: "Inside GitHub’s new Git storage architecture"
description: "GitHub announced Project Rubicon, a new storage architecture supporting thousands of writes per second and up to 5,000 agents per repo, tested at 7,200 pushes per second. The design minimizes coordination by decoupling storage from compute, leveraging that only reference updates need ordered agreement because Git objects are immutable, and indexing packs is CPU-heavy so writes use multiple nodes. Writes upload packs to Azure Blob Storage and commit a reference change to Ref-CAS, a log storing packs and references. Reads check the reference version with Ref-CAS and download missing packs into a virtual replica’s cache, which can be rebuilt from scratch. A compaction fleet merges packs and removes unreachable objects, and the private preview architecture, after migrating 99.99999% of repositories to Azure, aims to scale agentic development without changing Git client experience."
---

# Inside GitHub’s new Git storage architecture

[Github](https://yomu.fyi/company/github) · Karthik Raman · Nov 5, 2026

**Type:** Announcement

## Summary

GitHub announced Project Rubicon, a new storage architecture supporting thousands of writes per second and up to 5,000 agents per repo, tested at 7,200 pushes per second. The design minimizes coordination by decoupling storage from compute, leveraging that only reference updates need ordered agreement because Git objects are immutable, and indexing packs is CPU-heavy so writes use multiple nodes. Writes upload packs to Azure Blob Storage and commit a reference change to Ref-CAS, a log storing packs and references. Reads check the reference version with Ref-CAS and download missing packs into a virtual replica’s cache, which can be rebuilt from scratch. A compaction fleet merges packs and removes unreachable objects, and the private preview architecture, after migrating 99.99999% of repositories to Azure, aims to scale agentic development without changing Git client experience.

## Context

The rise of agentic development, where multiple AI agents work in parallel on a single repository, is driving exponential growth in write and read traffic, requiring a new storage architecture that can handle thousands of writes per second and scale reads on demand.

## Approach / What changed

Project Ubicon decouples storage from compute by using a fleet of disposable virtual replicas that cache data from Azure Blob Storage (for Git objects) and Ref-CAS (for references and pack manifests). Writes upload packs to Blob Storage and commit a reference change to Ref-CAS; reads verify the reference version and download missing packs into the local cache. A separate compaction fleet mergers packs in the background. This allows independent scaling of read, write, and compaction workloads.

## Takeaways

- The new architecture handles up to 7,200 pushes per second and supports 5,000 concurrent agents per repository, demonstrated in internal testing.
- Only reference updates require ordered agreement because Git objects are immutable and content-addressed; this allows parallel writes without coordination.
- Reads are served from a local disk cache on virtual replicas, which can be rebuilt from Blob Storage and Ref-CAS, and replica fleets can be scaled up as traffic grows.

## FAQ

### How does the new architecture handle high write throughput?

Writes upload packs to Azure Blob Storage and commit a single reference change to Ref-CAS. Indexing and connectivity checking are CPU-intensive, so work is spread across multiple writer nodes. Only the tiny reference update needs coordination, allowing parallel writes.

### What is the role of virtual replicas in the architecture?

Virtual replicas are disposable compute nodes that serve both reads and writes from a local disk cache. They pull data from Azure Blob Storage and Ref-CAS as needed, and their caches can be rebuilt from scratch. Spokes routes traffic to keep caches warm, and more replicas can be added for higher load.

### How does the architecture ensure reads are fast despite high write rates?

Reads check the reference version with Ref-CAS and download only missing packs into the local cache. A background compaction fleet merges packs and removes unreachable objects to keep reads efficient. Read amplification from writes is mitigated by caching and the ability to add more virtual replicas.

**Tags:** [Agents](https://yomu.fyi/topic/agents), [Architecture](https://yomu.fyi/topic/architecture), [Git](https://yomu.fyi/topic/git), [GitHub](https://yomu.fyi/topic/github), [Storage](https://yomu.fyi/topic/storage)

- Source: [Github](https://github.blog/engineering/architecture-optimization/inside-githubs-new-git-storage-architecture)
- Source URL: https://github.blog/engineering/architecture-optimization/inside-githubs-new-git-storage-architecture
- Ingested by Yomu: 2026-11-05T21:00:18.217Z

[Read original post](https://github.blog/engineering/architecture-optimization/inside-githubs-new-git-storage-architecture)
