Vibeleaderboard
← All Intel
Intel / post

Why agent-scale Git forces GitHub to rebuild its storage layer

Source
x.com
Date
GitHub@github

Git activity more than doubled between September 2025 and August 2026, reaching 473.3B events per month. We’re rebuilding our Git infrastructure while GitHub keeps running, creating a foundation for agent-scale software development. https://t.co/H5J2qxCqMT

Context

GitHub says that, for its busiest repositories, it can no longer absorb agent traffic simply by adding more servers, because in its current design the copies that keep repositories safe are also the copies that serve readers. In an engineering post by Brian Celenza, the company explains that a system called Spokes stores each repository as a full copy on the local disks of several servers, five by default. Every copy takes part in every write, so a push finishes only as fast as the slowest copy. Adding a copy to handle more clones and fetches therefore makes every push slower, and if too few copies remain to form a quorum, the minimum number that must agree on an update, writes stop entirely.

Agents push that tradeoff toward its limit. GitHub reports that pushes grew 4.9 times in a year, from 0.69 billion to 3.35 billion a month, and that the busiest repository saw roughly a billion requests in August 2026. The post describes an agent in a tight loop that commits after nearly every action, so the time a single push takes, which a person would never notice, becomes the agent's speed limit. Each push then fans out into reads: CI and code scanning fetch the same branch thousands of times a minute. GitHub says the current design still works well for most repositories; the ceiling appears at the highest activity levels.

The planned fix separates the two jobs. Authoritative repository data would live in Azure Blob Storage, while lightweight compute workers cache data to answer requests, so read capacity can grow without adding work to each push, and a failed worker behaves more like a cache miss than a lost replica. Within a push, only the reference update, the step that moves a branch to its new commit, would require agreement; most of the heavier work, such as storing objects and scanning, could run alongside other writes. GitHub says the new architecture reached up to 35 times higher write in its own internal benchmarks, and the post describes a rebuild still in progress rather than a finished system. GitHub's follow-up post in the series covers how it reworked the write .

Terms in this piece · Glossary
  • File — A named unit of stored data that keeps its contents while no program is using it.
  • Secret — A value such as a password, key or token that lets its holder act as someone; it must be kept out of code and shared files.
  • throughput — How much model output a provider can generate over time, commonly measured in output tokens per second.
  • Path — A description of a file or folder location, written using directory names and separators.
More from GitHub
Recommended reads
Comments

Checking sign-in…

Loading comments…