Cursor has detailed a new Git storage architecture designed to handle the increasing repository workloads created by AI coding agents.
Called Continuity, the system was developed to address scalability and reliability limitations Cursor says become increasingly apparent as repositories grow larger and automated agents generate more code, pull requests and CI activity.
Cursor developed the system as part of its work on Origin, its platform for hosting and operating Git repositories.
Rather than treating individual Git repositories stored on servers as the primary source of truth, Continuity uses a write-ahead log stored in S3-compatible object storage. Local Git repositories running on fast NVMe storage effectively operate as caches that can be recreated from the log when needed.
This approach allows repositories to scale across an arbitrary number of replicas without requiring every copy to participate in a consensus process for each push.
Continuity has been tested with as many as 100 replicas and says read performance continued to scale linearly without reducing push throughput.
In Cursor’s own synthetic testing, clusters using Amazon S3 Standard sustained up to 120 pushes per second, while deployments using the lower-latency S3 Express One Zone exceeded 300 pushes per second. Cursor says Git’s own on-disk compaction performance becomes the limiting factor at that level.
The architecture also allows lightly used repositories to operate with a single replica, or no active local copy at all until one is required, while heavily used monorepos can be distributed across substantially more machines.
Leave a Reply