GITBENCH

Real-time latency benchmarks for git hosting providers.
Measured from Vercel Sandbox (us-east-1) across 7 providers, 3 repo sizes, 7 operations.

CodeStorage

GitHub

GitLab

Codeberg

CF Artifacts

CodeCommit

CC us-west-1

Repo size:

create-next-app · 9 files

Clone

Full Clone

git clone (full history) — P50 on medium repos (ms)

Shallow Clone

git clone --depth 1 — P50 on medium repos (ms)

Push

Push 1 probe commit — P50 on medium repos (ms)

Revert

Force-push HEAD~1 to restore — P50 on medium repos (ms)

Fetch

ls-remote

git ls-remote (ref advertisement) — P50 on medium repos (ms)

Fetch (noop)

git fetch when up-to-date — P50 on medium repos (ms)

Fetch (delta)

git fetch after 1 new commit — P50 on medium repos (ms)

Detailed Results

Full Clone

Medium repo

git clone (full history) — ranked by P50 latency

#	Provider	n	P50	P95	Avg	StdDev	Min	Max
1	CodeStorage	20	637ms	901ms	711ms	±222ms	519ms	1,492ms
2	GitHub	20	642ms	857ms	680ms	±95ms	566ms	885ms
3	GitLab	20	889ms	1,140ms	926ms	±123ms	751ms	1,204ms
4	CF Artifacts	20	956ms	1,151ms	996ms	±116ms	802ms	1,207ms
5	Codeberg	20	1,021ms	1,155ms	1,045ms	±98ms	921ms	1,302ms
6	CodeCommit	20	2,996ms	3,104ms	2,987ms	±87ms	2,810ms	3,129ms
7	CC us-west-1	20	3,190ms	3,647ms	3,249ms	±196ms	3,091ms	3,854ms

Size Scaling

How latency scales from small (2 files) to large (10K files)

Full Clone

P50 by repo size

Push

P50 by repo size

Fetch (noop)

P50 by repo size

Fetch (delta)

P50 by repo size

Methodology

All benchmarks are run from a Vercel Sandbox in us-east-1. Each operation is timed end-to-end including network, TLS, and git protocol negotiation over authenticated HTTPS.

Test Repos

SmallREADME only · 2 files

Mediumcreate-next-app · 9 files

LargeSynthetic monorepo · 10,062 files

Operations Tested

Clone— Full clone and shallow clone (--depth 1) of each repo. All providers tested in parallel per round.

Push + Revert— Clone the repo, add a probe commit, push it, then force-push HEAD~1 to restore the original state. Both the push and the revert are timed separately.

Fetch— Three variants: ls-remote (ref advertisement), fetch when already up-to-date (noop), and fetch after a new commit was pushed (delta). The delta test pushes a probe commit, fetches from a separate clone, then reverts.

Clone and fetch benchmarks: 20 rounds. Push benchmarks: 15 rounds. Providers that return HTTP 429 are automatically disabled for remaining rounds (applies to Codeberg during clone benchmarks). Sandboxes that crash are automatically replaced and benchmarking resumes from the next round.