tobi/walgit

Source: https://github.com/tobi/walgit

walgit — a git server that is one binary in front of an object store

walgit hosts git repositories with no database, no leader and no local state that matters. You run a single binary, point it at an S3 or GCS bucket, and you have: smart HTTP (v0/v2) fetch and push, bundle-uri clones served as static files, Git LFS, a browsing web UI, a JSON API with an SDK, per-repository push policy, webhooks — and a server that scales to repositories larger than the machine it runs on. Every machine that runs walgit is a disposable cache; the bucket is the repository.

# 1. a bucket (any S3-compatible store or GCS) and a config
cat > walgit.toml <<'EOF'
[server]
listen = "0.0.0.0:8080"
public_url = "https://git.example.com"
auto_create_on_push = true
[server.auth]
mode = "token"
anonymous_read = false
tokens = [{ principal = "me", token_env = "WALGIT_TOKEN_ME", write = true }]
[store]
backend = "s3"
bucket = "my-walgit"
[store.s3]
endpoint = "https://s3.us-east-1.amazonaws.com"
region = "us-east-1"
EOF

# 2. run it
WALGIT_TOKEN_ME=$(openssl rand -hex 24) walgit serve --config walgit.toml

# 3. use it — a push to a new name creates the repository
git -c http.extraHeader="Authorization: Bearer $WALGIT_TOKEN_ME" push https://git.example.com/acme/app.git main

That is the whole deployment. Add more machines pointed at the same bucket and they serve the same repositories, consistently, with nothing to coordinate. Kill them all and you lose warmth, nothing else.

It is a Rust implementation of the architecture Cursor described in Git at any scale (the system they call Continuity), with the changes needed to run it on machines that are smaller than the repository. The post is worth reading first; it is kept verbatim in docs/reference/cursor-git-at-any-scale.md.

Why this shape

Git is distributed, and that makes hosting it miserable for one reason: packfiles. Everything in a repository is compressed into large binary packs laid out to be small, not to be read in order; every git operation is a random walk over gigabytes. That is fine on a laptop with the file in page cache and catastrophic over a network filesystem, which is why “just put the repositories on NFS” failed at every large host that tried it. The design that survived (GitHub’s Spokes) keeps real repositories on local NVMe so upstream git does the work, and replicates at the packfile level with strict consistency — paid for with three-phase commit across a fixed replica set, a database that maps every repository to its machines, and a fleet of pets.

Continuity’s insight changes the economics: make a write-ahead log in object storage the source of truth, and make every on-disk repository a cache. A push is stored as an immutable object in the bucket and becomes visible only when a tiny manifest is rewritten with a compare-and-swap. That CAS is the consensus — no election, no quorum, no primary. Any instance may accept a push; two racing instances cannot both win. A replica that has never seen a repository reads the log and has it. Reads are consistent without coordination because every read first asks the store whether anything changed (a conditional GET, usually a 304). Compaction is done once by whoever holds a lease and published into the log, so replicas download compacted packs instead of repacking. And because the WAL is the truth, there is complete provenance: every push and every repack, replayable to any point.

walgit takes that as-is, and adds what a monorepo on small machines needs: serving refs and web pages for a repository whose packs will never fit on the instance (a remote reader over HTTP range requests), keeping commits and trees local while blobs stay in the bucket (the history pack), and moving clone bytes out of the server entirely (bundle-uri: fresh clones and catch-ups are static files the bucket or a CDN hands out).

What it does

How it works, briefly

The repository is a WAL in the bucket. Under repos/<owner>/<repo>/: manifest.pb (tiny, CAS-rewritten: head sequence, the live pack set, checkpoint pointer, settings — the linearization point), log/<seq>.pb (immutable entries: PUSH, COMPACT, CHECKPOINT, SETTINGS), wal/<checksum>.pack|.idx|.rev|.bitmap|.commit-graph (immutable, content-addressed packs with their side-files), checkpoints/<seq>/ (folded ref snapshot + pack inventory so a cold start is snapshot + tail), bundles/, leases/ (CAS with TTL — the only cross-instance mutex), policy.json, lfs/objects/, events/cursor.json.

A push: our receive-pack indexes the pack (git index-pack --fix-thin --rev-index in a scratch dir), checks connectivity and policy, uploads pack ∥ idx ∥ log entry, then CASes the manifest. On a 412 it re-reads, re-validates every ref’s old value and retries. Concurrent pushes to one repository on one instance are group committed into one CAS. The client sees ok only after the bucket does.

A read: one conditional GET of the manifest; 304 → serve from the local copy, 200 → apply the new entries. What “apply” means depends on what the request needs: refs (snapshot + log → packed-refs, no packs: advertisements, the API, bundle lists), serve (the pack set as this machine can hold it: small packs and the history pack local, a too-large base read by range), full (everything local, for repacks), objects (the remote reader, for the UI on a repository that does not fit). Pack downloads run on their own runtime and never block a refs request.

Placement is configuration. [placement] serve / maintain globs say which repositories a host does object work for; refs-level reads work everywhere. One box: leave the defaults. Several: put the monorepo on the host with the SSD (cache.mode = "disk"), everything else on the small ones, and route by /<owner>/<repo> in front.

Nothing waits silently. Anything slow is a task with an id, a log and a progress stream — narrated to git on sideband 2 (remote: * …) and to the browser as SSE.

AGENTS.md is the full architecture and operating manual: constraints, the WAL strategies, every design decision with its reasoning, the invariants, and the cost model (round trips to the bucket are the budget).

Running it

# build (needs rust per rust-toolchain.toml, protoc, node 24 + pnpm for the web UI)
just web-build && cargo build --release -p walgit-cli
# or: nix build .#walgit        or: podman build -t walgit -f Containerfile .

# one box, TLS by walgit itself, a local S3 store (rustfs in a container)
just dev-store
./target/release/walgit-server --config walgit.standalone.toml
open https://walgit.localhost:8080/

• walgit.standalone.toml — the one-machine shape (self-signed TLS, rustfs, every role). Start here.
• walgit.example.toml — every key with its default and a comment.
• Containerfile, flake.nix — an OCI image and a Nix package/devshell.
• deploy/nginx.conf.example — an optional nginx in front: public TLS, one auth_request per credential, and byte offload: walgit answers bundle/LFS downloads with X-Accel-Redirect and nginx streams + caches the object from the bucket itself (S3 presigned or GCS with walgit’s bearer). The file documents the contract.

Roles (server.roles): serve (git, API, UI, bundles, LFS), maintain (checkpoints, bundles, compaction, fsck/repair), events (the webhook bridge). Empty = all. Any number of serve hosts may point at one bucket; give each repository one maintainer (placement globs) and you are done.

Authentication

Developer setup is one idempotent command — sh -c "$(curl -fsSL 'https://git.example.com/services/public/install.sh')" — which stores the token in a file only the user can read, installs a tiny git credential helper (git ≥ 2.46: it answers get with authtype=Bearer, and on a real 401 erases the token and says where a new one comes from), and turns on transfer.bundleURI. ?repo=owner/name clones right after.

Developing

just test          # fast hermetic tier (< 1 min): unit + quick integration, in-memory store, real git
just e2e           # real git against the server (~20 s)
just warnings      # zero rustc warnings across all targets
just ci            # all of the above
cargo test -p walgit-server --test sim     # fault-injection simulation (crashes, partitions, stale reads)
just test-s3       # store contract against local rustfs

Code map:

crates/
  walgit-proto    protobuf schema (wal.proto), log framing, store keys
  walgit-store    ObjectStore trait (CAS versions, conditional GET, range, compose); backends s3, gcs, memory; leases
  walgit-git      bare repos on disk, receive-pack, pack ingest, refs ↔ packed-refs, advertisements, upload-pack drivers
  walgit-wal      RepoHandle: sync levels, publish (group commit + CAS), checkpoints, log reader, remote reader, tasks
  walgit-bundle   bundle-uri: slots and chains, building, header ∘ pack composition, lists, retention
  walgit-server   axum: smart HTTP, LFS, bundles, auth (none/token/oidc), the maintainer loop, upstream follow,
                  web/ (API, UI, SDK routes, SSE), setup.rs (installer + recipes), events bridge
  walgit-config   walgit.toml (+ WALGIT__ env overrides), per-repo settings merge, fail-closed validation
  walgit-cli      `walgit serve|import|compact|bundle|wal|mirror|synth|config|repo`; `walgit-server` = `walgit serve`
web/              React SPA (Vite) + sdk/repos.ts, built into the binary; the wire contract is web/API.md
docs/             BUNDLE_URI_DESIGN, ROUNDTRIPS (the cost model), POLICY, LFS, INTEGRITY, EVENTS, CONTRACT, patches/

Invariants worth memorising

• The manifest CAS is the only commit point; everything before it is invisible, everything after it is idempotent and replayable.
• Immutable objects are content-addressed; nothing is overwritten except the manifest, the bundle list and leases.
• Every read revalidates against the bucket first; there is no “eventually”.
• Local disk is a cache. Memory is a cache. The bucket is the repository.
• Placement is configured, never inferred; refs-level reads work everywhere, object work only where placed.
• The maintainer’s output is a pure function of (config, WAL); missing is just “not built yet”.
• Cost must not scale with ref count on any hot path, nor with pack size on a machine too small for the pack.
• Long work is a task: discoverable, attachable, narrated.
• Correct is not sufficient: every protocol change is judged on round trips to the bucket (docs/ROUNDTRIPS.md).

License

MIT — see LICENSE.