Intel / article

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

Source: x.com
Author: GitHubTop Viber
Date: Published Oct 6, 2026

Context

GitHub says the way it stores repositories makes writes slower as it adds capacity for reads, and that this has become a ceiling for the busiest repositories. In an engineering post, the company explains that each repository is kept as a full copy on the local disks of several file servers, five by default, and every copy takes part in every write. A push therefore finishes only as fast as the slowest copy, so adding copies to serve more readers adds work to every writer.

Agents make that tradeoff easier to hit. GitHub reports that pushes grew 4.9 times in a year, from 0.69 billion to 3.35 billion a month, and describes an that saves its work after nearly every action as limited by how quickly a single push completes. The busiest repository saw roughly a billion requests in August 2026. These are GitHub's own figures, and the post says the current design works well for most repositories.

The post describes a rebuild still in progress, so it establishes why the storage layer is being changed, not yet how well the replacement performs.

Key takeaways · AI-distilled

GitHub says total Git activity more than doubled between September 2025 and August 2026, from 218.2 billion to 473.3 billion events a month, and developers and agents made 7.38 billion commits in September alone, over five times the year before.
In GitHub's current Spokes design, each repository lives as full copies on several fileservers (five by default) and every replica joins every write, so adding replicas to absorb reads makes pushes slower, and losing quorum stops writes entirely.
GitHub's redesign shrinks a push's coordinated critical path to the reference update itself: most of the work of storing objects, checking connectivity and secret scanning can run in parallel, and and garbage collection move to separate workers.
The new architecture keeps authoritative repository data in Azure Blob Storage and serves reads from cache-backed compute workers, so CI and agent fan-out scale without new durable copies. GitHub reports up to 35x higher write throughput in internal benchmarks.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.

Read the source x.com