GitHub plans to separate Git storage from serving compute

GitHub explains why more automated development puts pressure on its repository infrastructure. In the existing design, adding read replicas also adds participants to write coordination. The proposed replacement separates durable object storage from caching compute workers, narrows coordination to reference updates, and moves repository maintenance away from the machines serving interactive Git requests.

GitHub · Oct 6, 2026

Why it matters

The design illustrates how read scaling can accidentally make writes harder when replication and coordination are tightly coupled.

More context

Read the architecture discussion for its treatment of concurrent pushes, CI read fan-out and maintenance. GitHub is building the replacement around Azure Blob storage with independent workers; developers do not need to migrate their repositories to apply this design lesson.

Keep in mind

This is work in progress, not an announcement that the migration is complete. The post does not establish a universal latency improvement. Git object storage and reference consistency still have distinct correctness requirements.

Original read