I taught a bucket to speak Git

3 hours ago

The author developed objgit, a git server that stores repositories directly in object storage buckets, eliminating the need for local disk storage and the git binary.
By leveraging go-git and the billy filesystem abstraction, the author taught a Tigris bucket to mimic a filesystem, enabling git operations to function over object storage.
Git's storage model uses immutable, content-addressed objects (blobs, trees, commits, refs), making it compatible with append-only object storage systems like Tigris.
Traditional git servers rely on local filesystems, creating single points of failure; objgit decentralizes storage using object storage for improved resilience.
The author addressed challenges like implementing atomic renames (using Tigris's RenameObject extension) and reducing latency from excessive API calls (e.g., stat, ListObjectsV2).
Optimizations included caching pack files locally to speed up clones and modifying transport layers to handle packfiles efficiently, reducing round trips.
Post-receive hooks are supported in a sandboxed environment (using kefka), enabling automated workflows like deployments, though policy enforcement is not yet implemented.
Future plans include integrating CI/CD pipelines (e.g., Kubernetes, Tekton), adding a web UI, and improving performance and stability for production use.
Objgit is an experimental project, emphasizing its potential for stateless, scalable git hosting while cautioning against production use without further testing.

Hasty Briefsbeta