grit://cooking/dogfood← what’s cooking
Tuesday, November 10, 2026 · macOS 15.7 (24G214), Darwin 24.6.0

Dogfood on scott-mbp: Large-repository performance

4 repos · 39 commands · 2 broken · 3 slow · 3 filed

Dogfooded grit main (0.6.0, feb1e5b50) against system git 2.52.0 on git/git, rust-lang/cargo, Hello-World and a synthetic 60k-file tree: status, add, commit, log, diff, blame, revs, describe, switch/reset, stash, merge, rebase, fetch, clone (file:// and https), gc, fsck, plus --json/--markdown. grit matches or beats git on almost everything and produced identical trees and clean `git fsck`; the real gaps are blame speed (2x on git/git Makefile), describe without a commit-graph (known), and a few correctness drifts (blame attribution on ~1-2% of lines, diff hunk headers ignoring gitattributes drivers, log --follow missing 3 commits).

os
macOS 15.7 (24G214), Darwin 24.6.0 · arm64
machine
scott-mbp (agador.sparticus), Apple M3, 8 cores, 24 GB RAM
versions
grit 0.5.0 (~/.local/bin/grit; no blame/describe commands; fails to clone git/git) · main feb1e5b5 · 2.52.0
repositories4
repofilescommits.gitnotes
git/git—82,797—
rust-lang/cargo—23,703—
octocat/Hello-World—3—
cargo + 60,000 synthetic .rs files (copy)—23,703—
commandsgrit vs. system git, wall time
commandrepogritgitgrit/gitresultnotes
status (clean)git/git12 ms12 ms1.0×ok
status (clean)cargo25 ms26 ms1.0×ok
status (3000 touched+modified, 500 untracked)git/git20 ms16 ms1.3×ok
status (60k untracked files)cargo+60k90 ms240 ms0.4×ok
add (stage everything, 1506 files)git/git200 ms990 ms0.2×ok
commit (1506 files)git/git50 ms30 ms1.7×ok
add (stage everything, 1465 files)cargo310 ms730 ms0.4×ok
add (60k new files)cargo+60k5.08 s16.37 s0.3×ok
commit (60k new files)cargo+60k550 ms570 ms1.0×ok
revs HEAD --count / rev-list --count HEADgit/git305 ms374 ms0.8×ok
revs HEAD --countcargo83 ms98 ms0.8×ok
log -n 50 --statgit/git8.6 ms7.1 ms1.2×ok
log --oneline -- builtin/commit.c (731 commits)git/git310 ms320 ms1.0×ok
log --follow -- builtin/commit.cgit/git720 ms860 ms0.8×wrong
log --author/--grep/-Sgit/git390 ms510 ms0.8×ok
diff HEAD~1 HEADgit/git4.2 ms4.0 ms1.1×ok
diff --stat v2.0.0 v2.40.0git/git200 ms410 ms0.5×ok
diff --patch-text (2996 dirty files)git/git478 ms433 ms1.1×wrong
diff (default pretty)git/git260 ms390 ms0.7×ok
blame Makefilegit/git1.08 s532 ms2.0×slow
blame builtin/commit.cgit/git468 ms305 ms1.5×slow
blame read-cache.cgit/git648 ms602 ms1.1×ok
blame t/t0000-basic.shgit/git360 ms610 ms0.6×ok
blame Cargo.tomlcargo60 ms70 ms0.9×ok
describegit/git1.17 s21 ms55×slow
describecargo3.4 ms125 ms0.0×ok
switch --detach v2.0.0git/git710 ms870 ms0.8×ok
reset --hard mastergit/git520 ms410 ms1.3×ok
stash push + pop (3000 modified)git/git1.11 s2.09 s0.5×ok
merge (fast-forward v2.50.0..v2.52.0)git/git210 ms550 ms0.4×ok
rebase 30 commits onto v2.50.0git/git1.08 s1.98 s0.5×ok
fsckgit/git8.95 s25.93 s0.3×ok
fsckcargo820 ms4.25 s0.2×ok
gccargo1.02 s1.70 s0.6×ok
clone httpsgit/git14.90 s21.01 s0.7×ok
clone httpscargo5.11 s5.24 s1.0×ok
clone file://git/git9.98 s16.02 s0.6×ok
fetch 3000 new commits (file://)git/git3.71 s9.11 s0.4×ok
status --json / diff --json / log --jsongit/git110 ms——ok
worked5
  • ›Trees produced by grit add/commit/rebase/stash are byte-identical to git's; git fsck --strict is clean on every grit-written repo including after grit gc (commit-graph verify OK).
  • ›status, add, commit, switch, reset, stash, merge, rebase, fsck, fetch and clone are on par with or faster than git on git/git and cargo, including a 60k-untracked-file tree (add 3x faster).
  • ›Log/revs filters (path, author, grep, -S, -G, merges, ranges, --all) give exactly git's commit counts.
  • ›--json outputs parse and are bounded; log paginates with a cursor.
  • ›grit clone of git/git over https works on main in 15 s.
problems9
  • ›blame is 2x slower than git on git/git Makefile (filed #2572; note: I wrongly wrote in that issue that release 0.5.0 behaves the same; 0.5.0 has no blame command).
  • ›diff hunk headers ignore gitattributes diff drivers (filed #2573; the 'not seen in 0.5.0' remark there is unverified).
  • ›blame attributes ~1-2% of lines differently from git on git/git (filed #2574).
  • ›describe takes 1.17 s / 134 MB vs git 21 ms / 26 MB on git/git without a commit-graph, 50 ms vs 9 ms with one: not filed, same root problem as open #2462/#2467 (numbers improved from 46 s).
  • ›log --follow omits 3 commits for builtin/commit.c (605 vs 608), 2 for read-cache.c: not filed, duplicate of open #2459.
  • ›Release 0.5.0 on PATH cannot clone git/git (I/O error: failed to fill whole buffer) and lacks blame/describe: release-only, works on main.
  • ›grit clone max RSS is 4x git's on git/git (444 vs 112 MB) but macOS peak memory footprint is 74 vs 58 MB: not filed, mmap RSS artifact (closed #2410). grit clone file:// of git/git peaks at 1.0 GB RSS vs 685 MB for git (1.47x, below the bar; footprint not measured).
  • ›UX nits not filed: `grit revs --count` errors 'at least one revision range is required' instead of defaulting to HEAD; `grit switch -q` is rejected; `grit status -s` prints '.M' where git prints ' M'.
  • ›git add -A baseline on git/git was index-stat-cold (my git copy had not run status first), so the 5x add gap is partly unfair; the 60k-file add (16 s vs 5 s) is like-for-like.
bugs filed3
#2572 #2573 #2574already reported: #2459 #2462 #2467 #2410
next time
  • ›Re-test describe and blame after a commit-graph write with changed-path filters (blame -L, log -- path).
  • ›Try a 1M-file monorepo-style tree, a 500k-commit synthetic history, and a repo with large binary blobs.
  • ›Rebase with conflicts (--continue/--abort) and history ops (reword/squash) on git/git; grit push to a bare repo.
  • ›Measure macOS peak memory footprint instead of max RSS everywhere, since mmapped packs inflate RSS.
  • ›Run on Linux and Windows.