What makes a Git GUI slow on large repositories?
See what MoonGit measured on million-commit history and a 500,000-file working tree, why each workload behaved differently, and how to compare your own.
Two repositories can feel slow for completely different reasons. One may have a deep commit graph. Another may have a relatively short history and hundreds of thousands of working-tree files. Measuring only how long it takes to open a window can hide the difference.
In September 2026 we measured both cases in MoonGit using fixed test repositories. We changed when the app does certain work and measured the same service operations again.
What we measured
| Workload | Before | After |
|---|---|---|
| Open a history session and lay out its first 250 rows, synthetic one-million-commit history, warm median | 10.32 s | 4.90 s |
| Open basic status and history, 500,000 tracked files with 100,000 modified and 100,000 untracked, fresh-process median | 125.16 s | 15.51 s |
| Aggregate analytics over one million commits, warm median | 14.36 s | 9.08 s |
These are historical before-and-after timings for MoonGit service operations on the same Windows development host. They measure specific operations inside the app. Desktop launch, input latency, network transfers, and other Git clients were outside these measurements. The host used Windows x64, 64 GiB of RAM, 28 logical processors, .NET 10, and Git for Windows. A complete distribution included five fresh-process samples and 20 recorded warm samples after warm-up. A fresh process could still benefit from Windows’ filesystem cache. The synthetic linear history was designed to isolate traversal cost. Real merge-heavy histories can behave differently.
A big graph makes repeated work expensive
The graph used to open the repository and read the first page, then initialize another topological walk for older rows. On a linear history of a million commits, initializing the walk twice meant paying for the same work twice. We gave the first page and its continuation the same repository and enumerator. The original comparison kept the number of rows, graph roots, and ordering the same. A later bounded cache can make a repeat open through the same service much faster, so a cached reopen belongs in its own measurement.

A busy working tree makes optional details expensive
Opening a tree with 200,000 modified or untracked paths used to calculate line statistics before returning basic status. We moved those calculations to when a visible row needs them. Users can still see whether a file was changed, and the app can distinguish a pending count from a known zero or an unavailable count. Managed allocations for the same fresh-process workload fell from 2.30 GB to 150 MB. Those are bytes allocated by the managed runtime during the operation, a different measurement from peak process memory.
Analytics benefited from a different change. We replaced retaining full commit records with incremental totals grouped by the report’s categories. That preserved the results while reducing how much data stayed in memory.
There were still expensive cases. A separate full Chromium checkout with 507,889 tracked files took a median of 61.21 seconds to open in a fresh process. The sample exceeded the campaign’s proposed opening target. It is a reminder to keep real checkout sizes alongside synthetic ones and to report first and repeat opens separately.
Compare Git clients on the repository you use
Record the parts of the repository that make your case expensive. These read-only PowerShell commands describe commits reachable from your current tip, the tracked index paths, and the current working tree’s changes.
git rev-list --count HEAD
@(git ls-files).Count
git status --porcelain=v1
The path list and status can be large. Run these before the timed trials so their work doesn’t get included. Warm storage can still affect the trial.
- Test the same checkout and Git state. Record the current commit and list any local changes. If your app calculates optional details, record those settings.
- Decide what counts as ready. For example, measure when you can select a graph row, then time opening one changed file as its own action.
- Repeat from a fresh process and in an already-running app. Keep the two groups separate and record the median and any unusually slow runs.
Commit count alone misses a busy index, thousands of refs, and expensive diffs. Use the measurements to isolate the operation you actually want to improve. See how to narrow MoonGit’s commit history and how sparse checkout limits working-tree paths.
MoonGit