Portfolio project — git infrastructure

Git object storage.
Three backends.
One verdict.

A working git remote server in Go with swappable storage backends. Built to empirically reproduce the tradeoffs teams hit when building git infrastructure for AI coding platforms.

See the numbers Run live benchmark
$ git clone https://git.wyat.me/git-storage.git
Repo creation throughput
git init --bare · 16 workers · single machine
This server local, bare metal
16,212 repos/min
baseline
Pierre sustained peak, cluster
15,000 repos/min
their stated peak
GitHub global avg
230 repos/min
70.5× faster than GitHub
GitHub API — p50
864ms
per repo creation
GitHub API — p95
1.06s
per repo creation
GitHub API — p99
2.253s
per repo creation
GitHub API latency measured directly via go run ./cmd/repobench -mode=github · throughput capped by secondary rate limits (~50 repos/min per account) · Pierre announced >15k repos/min sustained peak (Mar 2025) · GitHub announced ~230 repos/min global avg (Oct 2023) · This server measured locally via -mode=direct
Key findings
The numbers are not subtle.
2,548×
Exists — BadgerDB vs MinIO/S3 in production
Git calls Exists constantly during push to avoid resending objects. BadgerDB handles 395k ops/sec on Railway. MinIO/S3 handles 155 — every check is a network round trip to object storage regardless of object size.
20×
Concurrent Put — BadgerDB vs SQLite (1KB)
Without WAL mode and a connection limit of 1, SQLite produces SQLITE_BUSY errors under concurrent load. With those fixes, BadgerDB still wins by 20× at small object sizes in production.
39ms
MinIO/S3 Get p50 — in production, any size
A 1KB Get on MinIO costs 39ms in production. A 1MB Get costs roughly the same. The network round trip to object storage is the bottleneck, not the data. Git's workload is small objects.
~1×
Large object Put — all backends converge
At 1MB, SQLite (49), BadgerDB (52), and MinIO (47) are within noise of each other. The bottleneck becomes I/O, not the storage layer. Backend choice matters most at small sizes.
Benchmark results
ops/sec — higher is better
Loading benchmark data...
Put
Get
Exists
Concurrent Put
SQLite
BadgerDB
MinIO/S3
Operation Size SQLite BadgerDB MinIO/S3