Hi,

I'm working at switching Forgejo's implementation of getting the last modified commits in a directory to git-last-modified(1). I'd expected equal or better performance than the current implementation, but have not yet been able to get this and I'm a bit puzzled as to why.

The current implementation of Forgejo (inherited from Gitea) works
roughly like this:
1. Run `git log --name-status -c --format=commit%x00%H %P%x00" --parents
--no-renames -t -z $OID -- :(literal)some/path`, the output of this is
quite complex and possible outputs more information than necessary.
2. The output of this is piped to some code to a parser and reconstructs
what commit ID last modified each file in the directory.
3. Via `git cat-file --batch` get each unique commits information.

With git-last-modified(1) (-z --show-trees --max-depth=0) this replaces step 1-2, but is slower. I've isolated the degraded performance to the fact that git-last-changed(1) takes more time to finish. So from my perspective it does not seem worth it to replace the current implementation with git-last-modified(1), and I would like to know if I'm missing something here or if git-last-modified(1) possibly could see a speedup?

The repository I'm currently using to evaluate the performance is https://codeberg.org/ziglang/zig

Reproduction steps:
1. `git clone https://codeberg.org/ziglang/zig $(mktemp -d)`
2. cd to tmp directory.
3. `git commit-graph write --changed-paths`. As git-last-modified(1)
makes good use of the bloom filters.
4. `hyperfine 'git last-modified -z -t --max-depth=0
80d06578ac66bce3aa0a21e9610cdb782b9a0593 -- doc/langref/' 'git log
--name-status -c "--format=commit%x00%H %P%x00" --parents --no-renames
-t -z 80d06578ac66bce3aa0a21e9610cdb782b9a0593 -- ":(literal)doc/langref"'`

With as output:
Benchmark 1: git last-modified -z -t --max-depth=0
80d06578ac66bce3aa0a21e9610cdb782b9a0593 -- doc/langref/
 Time (mean ± σ): 66.5 ms ± 0.6 ms [User: 60.6 ms, System: 5.2 ms]
 Range (min … max): 65.3 ms … 67.7 ms 44 runs

Benchmark 2: git log --name-status -c "--format=commit%x00%H %P%x00"
--parents --no-renames -t -z 80d06578ac66bce3aa0a21e9610cdb782b9a0593 --
":(literal)doc/langref"
 Time (mean ± σ): 26.2 ms ± 1.0 ms [User: 17.3 ms, System: 8.4 ms]
 Range (min … max): 24.3 ms … 30.1 ms 110 runs

Summary
 git log --name-status -c "--format=commit%x00%H %P%x00" --parents
--no-renames -t -z 80d06578ac66bce3aa0a21e9610cdb782b9a0593 --
":(literal)doc/langref" ran
 2.54 ± 0.10 times faster than git last-modified -z -t --max-depth=0
80d06578ac66bce3aa0a21e9610cdb782b9a0593 -- doc/langref/

Kind Regards Gusted

On Tue, Jul 14, 2026 at 08:33:59PM +0200, Gusted wrote:

Thanks for this concrete reproduction. I can see the same problem here. Interestingly, if we turn off changed-paths, we get very different results.

Without a commit graph at all, last-modified wins (this is using the zig repo and the commands above):

  - log: 150ms
  - last-modified: 79ms

But with a graph and no changed-paths, they're about equal:

  - log: 61ms
  - last-modified: 61ms

And then with changed-paths, the log command gets much faster but last-modified gets slower!

  - log: 20ms
  - last-modified: 64ms

I think there's a tradeoff in the way that last-modified uses the bloom filters. It makes a key for every path we're interested in, and then for each commit, we check each key to say "is this in the commit's filter?".

So if you have a subdirectory with a non-trivial number of entries (like doc/langref here which has 290), but most commits don't touch that path at all (only 120 out of ~39k in this case), we'll spend a lot of time checking each key against each filter. We save ourselves opening the trees, but at the cost of 290*39k filter comparisons).

Whereas in the git-log case, we make a filter key out of the single pathspec we're given, and then check each commit against that. So we only do a single filter check for each commit to narrow it down to those 120 that matter (modulo a few filter false positives).

But I don't see any reason that last-modified couldn't _also_ do that: pre-filter the commits with a commit matching the original pathspec, and discard most commits with a single filter check.

The hacky patch below does this, and brings my last-modified runtime down to 16ms (a 4x improvement, and just a bit faster than git-log).

It tries to reuse the logic from revision.c, so it's doing the exact same filtering that git-log would do. I think there are other ways to do it. E.g., we could make our own "root" bloom key that contains all of the paths and pre-filter with that. But it seemed to be a little slower when I tried it (~24ms). I'd guess that the problem is that because the bloom filter is probabilistic, if you shove too many items into a single key you'll end getting more and more false positives. So putting all 290 entries into one key is too much, and we are better off just considering the shared prefix.

Anyway, here's the patch. Toon, I'm not planning to take it further immediately, but you may be interested in poking at it. It probably needs at least:

  - some light refactoring of revision.c

  - tests? We don't seem to cover last-modified with changed-paths at
    all, and just rely on the test-vars CI job which sets
    GIT_TEST_COMMIT_GRAPH_CHANGED_PATHS. It did pass for me with that
    flag, so surely I didn't introduce any bugs. :)

  - more timing exploration; e.g., might it make things worse if
    doc/langref were touched in 99% of the commits? Probably not, but it
    might be nice to check timings against a few repo shapes and request
    depths.

  - Not all pathspecs can support bloom filters (e.g., "*.c" would not).
    So in theory:

       git last-modified HEAD -- "*.c"

    could work, but wouldn't be optimized. I don't think it _does_ work
    now, because last-modified's max-depth logic complains. So it might
    be a non-issue.

    But I think it is solvable if we really wanted. Rather than
    traversing looking for "*.c", we actually expand the pathspec in the
    tip commit to a set of literal paths, and then as we traverse we
    look for those paths. So we could collect all of "*.c" and then
    add bloom keys for the shared prefixes. I think this does get tricky
    in the general case, though. If you have "a/b/c" and "a/b/d",
    looking for "a/b" is reasonable. But what if you also have "a/e"?
    Should you just have a key for "a/", or both "a/b" and "a/e"?
    There are some tradeoffs between how often uninteresting things in
    "a/" will give us a false positive, versus the cost of checking
    extra keys.

    So maybe an interesting area, but given that in practice most people
    will feed a single pathspec to last-modified, it's a lot easier to
    just use that.

  - I know that last-modified was derived from GitHub's blame-tree
    implementation (which I originally wrote, but stopped paying
    attention to well before it learned about changed-path filters). I
    don't know if the problem was solved separately there, but it would
    be worth checking. +cc Taylor

-Peff

---

Gusted <gusted@codeberg.org> writes:

Hi Gusted,

Thanks for reaching out.

You're actually not the first to notice this, and I've been aware of this.

The thing is, you're testing the difference on a single file. For us at GitLab, it wasn't very useful to optimize that use-case, because usually we want to see the last commit for a bunch of files at once. So the use-case for git-last-modified(1) for us has been to replace (pseudo code):

$ FILES=$(git ls-tree $COMMIT $PATH) $ foreach $FILE in $FILES; do git log -1 $COMMIT -- $FILE; end

GitLab is batching files 25 at once, and in my benchmarking, it was shown git-last-modified(1) is faster:

$ git last-modified $COMMIT -- <files

(I did this benchmarking in our Gitaly component to have a real-world experience and you can visit the results at: https://gitlab.com/gitlab-org/gitaly/-/merge_requests/7999#note_2850505479 )

So we left the door open for future improvement, although I never have gotten to it. At some point I was trying to chase down when git-log(1) was doing differently, but I never figured it out.

But this email challenged me already. And with some help of AI, I managed to work on some improvements. You can expect a patch series soon.

(Right before sending out this mail I noticed Peff sent out some changes as well. I'll coordinate how to combine.)

Jeff King <peff@peff.net> writes:

As I mentioned, I was aware of that issue, but never felt the need (and didn't have the deep Bloom filter knowledge) to fix it.

Oh, that's very useful of you to explain this. Thank you.

Funny, I was toying around with my AI agent, and they came with a similar solution, but I'm working on a cleaner solution.

Agreed.

Fair of you calling that out. Thanks for checking.

Maybe, I tried a few things.

On gitlab-org/gitlab (our Rails monolith), there seems to be a noticable improvement when running for `app/`:

    Benchmark 1: old
      Time (mean ± σ):     435.5 ms ±   9.8 ms    [User: 369.5 ms, System: 64.4 ms]
      Range (min … max):   425.3 ms … 450.9 ms    5 runs
     
    Benchmark 2: new
      Time (mean ± σ):     278.5 ms ±  32.3 ms    [User: 208.0 ms, System: 69.3 ms]
      Range (min … max):   246.2 ms … 314.8 ms    5 runs
     
    Summary
      new ran
        1.56 ± 0.18 times faster than old

The app/ directory is touched by roughly 35% of the commits.

In gitlab-org/gitaly, I ran a benchmark on `internal/`, which is touched in about 58% of the commits:

    Benchmark 1: old
      Time (mean ± σ):      13.2 ms ±   1.3 ms    [User: 10.4 ms, System: 2.5 ms]
      Range (min … max):    11.3 ms …  14.8 ms    10 runs
     
    Benchmark 2: new
      Time (mean ± σ):       9.9 ms ±   0.4 ms    [User: 7.3 ms, System: 2.4 ms]
      Range (min … max):     9.6 ms …  10.8 ms    10 runs
     
    Summary
      new ran
        1.33 ± 0.14 times faster than old

(although Gitaly a lot less commits, ~23k commits vs ~530k in our Rails monolith)

And also --recursive it's faster:

    Benchmark 1: old
      Time (mean ± σ):     204.6 ms ±   4.4 ms    [User: 193.2 ms, System: 10.5 ms]
      Range (min … max):   199.9 ms … 213.2 ms    10 runs
     
    Benchmark 2: new
      Time (mean ± σ):     181.5 ms ±   2.7 ms    [User: 170.9 ms, System: 9.7 ms]
      Range (min … max):   177.0 ms … 187.1 ms    10 runs
     
    Summary
      new ran
        1.13 ± 0.03 times faster than old

Personally I'm not too worried any use-case would be at least equally fast.

Yeah, I rather not deal with that right now.

I assume you're talking about check_maybe_different_in_bloom_filter()?

I was working on a fix to simply make it public and call it, but that's a very valid point you're making. I'll change my plans.

On 7/16/26 11:26 AM, Toon Claes wrote:

Hi Toon and Jeff,

Thanks for having a look at this!

The use case for Forgejo is the same as Gitlab then, we only use it to get the last-modified of each entry in a directory. Looking a bit closer I missed that in Forgejo's code it's considered the output might not have the answer of all entries, and happily calls it as many time is needed on the 'remaining paths', especially when using gitlab as repository this required a lot of git-log calls.

Looking at the mentioned benchmark, yeah this is where Forgejo's implementation with git-log would fail in terms of performance (seemingly even slower than doing N-1 git-log calls, by using your benchmark numbers). It's somewhere in the minutes :')

Kind Regards Gusted

On Thu, Jul 16, 2026 at 11:26:25AM +0200, Toon Claes wrote:

That was my assumption at first, too, but I think the log command there really is returning results for the whole subtree. You just have to post-process it to pick out the files from each commit.

Yeah, that is the most horrible way to do it. It's expensive in processes, but also in walking over the same set of history repeatedly.

The log in Gusted's example does a single walk, but it is up to the caller to then interpret the walk results. That would add extra time, but I think it scales independently of the time difference he's observing. In his hyperfine results, last-modified is scaling with the total numbers of commits in the repo, but processing the output scales to the number of commits which actually touched the subtree in question.

So I think it really could perform better than last-modified, even with the post-processing step (which we didn't see nor time). But we should be able to do better in last-modified using similar top-level commit filtering.

-Peff

On Thu, Jul 16, 2026 at 01:42:04PM +0200, Toon Claes wrote:

Yeah, I would be surprised to find a practical case where it makes things slower. Checking one bloom key is cheap-ish, and unless the subtree being queried is touched by almost every commit, it's going to be a net win.

So yeah, that's my gut feeling, too.

Yeah, exactly. We already do the first half (getting the commit's filter) ourselves. And then most of the rest is just trace2 accounting, which we don't necessarily need to do. So we're left with just that one bloom over the keyvecs, which is fairly trivial. Mostly it felt weird to be looking at the innards of rev_info, and the logic for what those keyvecs means should remain in revision.c.

I was just thinking to split it into two (get the filter, and then check the filter against the rev_info) and make the latter half public.

-Peff

Toon Claes <toon@iotcl.com> writes:

Hah, wait, ... the previous reporter was ... you[1].

I was mixing up use-cases, and the problem you're stating here isn't related to your previous report. Now while I did sent out patches for the problem in this email, the issue in the other email isn't addressed with it. Unfortunately.

(nerdsniping Peff :-))

Gusted <gusted@codeberg.org> writes:

That's quite clever actually.

My only worry would be this could end up in a very long list of (duplicate) commits. But you can probably filter out data as you read in lines.

Yes we use git-cat-file(1) in batch mode too.

But, and I learned this from blog post[2] from GitHub about Git v2.55, git v2.55 now has git-format-rev(1)[3]. You can pipe the output of git-last-modified(1) into that and get everything you want at once (maybe).

I haven't tried yet to integrate that into GitLab to see if it would bring any gains.

[1]: https://lore.kernel.org/git/03f96860-29fc-42a7-a220-c3ec65eb8516@codeberg.org/ [2]: https://github.blog/open-source/git/highlights-from-git-2-55/ [3]: https://git-scm.com/docs/git-format-rev