GitHub is overhauling its entire Git storage system from the ground up after internal testing revealed a 35-fold improvement in write performance, according to a blog post by Brian Celenza, a principal software engineer at the platform. The redesign responds to surging demand from AI coding agents that has doubled the company's traffic in a single year. The new system decouples read operations from write operations, shifts maintenance tasks to background processes, and relies on object storage to handle redundancy, all while aiming to preserve existing developer workflows without disruption.
Between September 2025 and August 2026, GitHub's monthly traffic climbed from 218.2 billion events to 473.3 billion, a twofold increase driven by the shift toward AI agent-based coding. Commits alone reached 7.38 billion in September 2026, five times the volume recorded in September 2025. This surge in activity triggered a wave of performance problems: the platform logged 10 incidents that degraded service in April, followed by nine more in May, each attracting broader negative attention than the one before.
GitHub's existing storage system, known as Spokes, relies on a three-phase commit protocol that stores complete repository copies across multiple local disks and requires a quorum of replicas to confirm each write. While this design delivers strong reliability, it creates a bottleneck for write speed because every push is limited by the slowest replica needed to reach quorum. Celenza wrote that the platform must now support "large engineering teams running busy CI pipelines alongside growing fleets of agents" and build Git infrastructure for sustained, concurrent reads and writes at a scale few repositories encounter today. He did not provide a timeline for completing the migration.
The replacement architecture writes each commit just once to Azure Blob Storage, an object store service that automatically manages replication and redundancy on its own, eliminating the quorum bottleneck that slowed the older system. Read requests are separated into their own channel and handled by lightweight compute workers, with only the reference branch pointers requiring coordination between reads and writes. Maintenance duties such as compaction and garbage collection are moved off the serving path and run as background processes, further cutting system lag. GitHub isn't alone in rethinking Git for the era of AI agents: former GitHub CEO Thomas Dohmke launched a competing service called Entire that offloads agent traffic to mirror repositories, while SpaceX subsidiary Cursor rejected the three-phase commit in favor of uploading pushes to an object storage write-ahead log that captures all changes as immutable objects and caches at least one copy on solid-state disks.
The report notes that GitHub engineers face a challenge few could have predicted when they designed the service, with one industry observer remarking that companies typically spend three months preparing for traffic increases of just 10 to 20 percent, a fraction of what GitHub now confronts. The new storage architecture is being deployed with the goal of avoiding any disruption to developer workflows, review processes, or security controls, and the hope is that it will provide GitHub's engineers with some well-deserved relief. The platform's ability to manage infrastructure at this velocity may determine whether it can maintain its position as the default collaboration layer for an industry increasingly powered by autonomous code generation. Organizations betting on agentic development workflows will likely scrutinize GitHub's uptime figures closely, knowing that any prolonged outage could ripple across entire software supply chains.