five git implementations, one repository, the same six operations.

Every runner is held to the same work contract on a full clone of vercel/next.js, so the numbers compare implementations rather than workloads. bun:ffi + libgit2 finishes 4.2× faster than the git CLI, and 356× separates the fastest runner from the slowest.

Read the methodology Run it yourself

5 runners · 6 operations · 5 timed samples each · last run Aug 24, 2026

Overall ranking

Geometric mean of the six operations against the git CLI, on a logarithmic axis: lower is faster, a bar left of parity beats git, and the whisker spans the runner’s best and worst operation.

RunnerScore

Which library should you use?

Recommendations are derived from the published results and update when the benchmark changes.

Best overall
bun:ffi + libgit2
Lowest geometric mean across all six operations, normalized against the git CLI.
Best in-process option
bun:ffi + libgit2
Avoids subprocess startup when the benchmark work is embedded in your application.
Most reliable
gitoxide
Fewest parity mismatches and failed operations, with the smallest average sample spread.

Operations

How far apart the implementations land on each operation: the bar is the slowest median divided by the fastest, and the whisker is that ratio at its most and least favourable across the samples.

OperationSpread

Log (100)Walk the 100 most recent commits from HEAD.12,785× ±24%0.26ms
Read 25 blobsRead 25 file blobs at HEAD by path.7,802× ±9%10ms
Changed filesName-only diff between HEAD~1 and HEAD.4,226× ±5%2.33ms
Tracked filesEnumerate every path tracked in the index.210× ±49%1.30ms
Current branchResolve HEAD to a branch name.149× ±62%0.04ms
StatusWorktree + index status (porcelain).21× ±5%364ms

See every runner and number

Speed against consistency

Score, lower is faster, against how far the same runner drifts between its best and worst operation. Both axes logarithmic; the brighter lines mark parity with git.

Score:: the geometric mean of the six operations, each as a multiple of the git CLI.
Spread:: the worst operation divided by the best. A point far to the right hides at least one operation you would not want in a request path.

Full results

Median time per operation; sort by any runner to find what it handles worst.

Operation
Current branch Resolve HEAD to a branch name.	1.47ms1.0×1.37ms–1.63ms	0.04ms0.03×0.03ms–0.06ms	5.90ms4.0×5.73ms–7.11ms	3.64ms2.5×3.20ms–6.02ms	1.49ms1.0×1.31ms–2.58ms
Status Worktree + index status (porcelain).	407ms1.0×405ms–421ms	687ms1.7×685ms–688ms	364ms0.89×350ms–377ms	7.59s19×7.44s–7.67s	1.85s4.5×1.83s–1.85s
Log (100) Walk the 100 most recent commits from HEAD.	5.07ms1.0×4.10ms–5.94ms	0.26ms0.05×0.22ms–0.35ms	12ms2.3×12ms–12ms	3.32s655×3.28s–3.36s	11ms2.1×10ms–11ms
Tracked files Enumerate every path tracked in the index.	17ms1.0×17ms–19ms	1.30ms0.08×0.90ms–2.66ms	273ms16×269ms–274ms	36ms2.1×34ms–50ms	16ms0.92×15ms–16ms
Changed files Name-only diff between HEAD~1 and HEAD.	2.33ms1.0×2.24ms–2.41ms	14ms5.8×13ms–14ms	30ms13×29ms–37ms	9.83s4,226×9.69s–9.88s	3.11s1,338×3.11s–3.12s
Read 25 blobs Read 25 file blobs at HEAD by path.	59ms1.0×57ms–64ms	10ms0.18×9.56ms–11ms	155ms2.6×151ms–165ms	81.45s1,377×79.53s–82.43s	64ms1.1×64ms–65ms

Methodology

Every runner executes in the same Docker image against one clone of vercel/next.js: one warmup iteration is discarded, 5 timed samples follow, and the median is published.
Runners are held to the same work contract — status collects untracked but not ignored files, blob reads materialize full contents — and any disagreement with the git CLI baseline is logged as a parity mismatch rather than counted as a win.
The score is the geometric mean of each operation's ratio to the git CLI, so no single slow operation decides the aggregate.
Numbers come from a GitHub Actions runner, so the comparison between runners is the signal and the milliseconds are context.

CPU: AMD EPYC 7763 64-Core Processor
Runtime: Bun 1.4.0
Source: github-actions

Run Aug 24, 2026 against 8057603. Raw results, config schema and the source.