git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH v2] describe: limit default ref iteration to tags

From
Tamir Duberstein <tamird@gmail.com>
Date
Jun 9, 2026, 14:44 UTC
Message-ID
<CAJ-ks9kz5JGFSF21aOhuXfgsJ+5aa5xE69RPT2Vhn-CRGyHZ6A@mail.gmail.com>
In-Reply-To
<20260609110957.GB1509396@coredump.intra.peff.net>
On Tue, Jun 9, 2026 at 4:09 AM Jeff King <peff@peff.net> wrote:
Show 34 quoted lines
>
> On Mon, Jun 08, 2026 at 07:32:14PM -0700, Tamir Duberstein wrote:
>
> > The benchmark checkout had 120,532 refs, of which 330 were tags. With
> > `$repo` naming the checkout, `$commit` an exactly tagged commit, and
> > `$parent` and `$this` the two binaries, I ran:
> >
> >     hyperfine --warmup 3 --runs 15 \
> >         --command-name parent \
> >         '$parent -C $repo describe --exact-match $commit' \
> >         --command-name 'this commit' \
> >         '$this -C $repo describe --exact-match $commit'
> >
> > The results were:
> >
> >     Benchmark 1: parent
> >       Time (mean ± σ):     171.7 ms ±  18.5 ms    [User: 23.9 ms, System: 133.6 ms]
> >       Range (min … max):   142.3 ms … 198.3 ms    15 runs
> >
> >     Benchmark 2: this commit
> >       Time (mean ± σ):       9.9 ms ±   1.1 ms    [User: 3.3 ms, System: 4.7 ms]
> >       Range (min … max):     8.8 ms …  13.1 ms    15 runs
> >
> >     Summary
> >       this commit ran
> >        17.35 ± 2.63 times faster than parent
> >
> > Both revisions were built with -O3, -mcpu=native, and ThinLTO using
> > Apple clang 21.0.0 on macOS 26.5. The machine was a MacBook Pro
> > (Mac16,6) with a 16-core Apple M4 Max (12 performance and four
> > efficiency cores) and 128 GB RAM.
>
> This patch looks fine to me, but let me pick a nit for a minute, because
> I think there is a broader conversation to be had.
Just to say from the start: I appreciate you taking the time to discuss this.
Show 10 quoted lines
>
> Given the discussion in earlier rounds and sibling topics, I assume the
> commit message here was AI-generated. And it's OK in the sense that it
> is describing what happened and I assume is entirely accurate. But as a
> human reader, it feels so much more verbose than what I'd expect, as it
> is full of semi-irrelevant details. Why set --warmup and --runs? Why
> bother with --command-name, which just means you have to show the
> commands separately anyway? Is the amount of RAM in the machine
> important for this test? Surely it could be if it was absurdly tiny, but
> in general, no, I would not expect it to be.

Well, the details matter in case some human reader knows something I don't, or wants to reproduce the findings and observes something completely different - they aught to be able to reconstruct my environment.

The command-name flag is needed; without it the output of the hyperfine would include local paths and would require post-processing to include in the message.

Show 13 quoted lines
>
> So while it is perhaps reasonable to document every detail in case
> somebody later wants to verify or reproduce timings, it is a little
> overwhelming when trying to tell a story, the core of which is:
>
>   In a repo with ~120k refs, ~300 of which were tags, running:
>
>     git describe --exact-match $some_tag
>
>   went from ~170ms to ~10ms, since we no longer needed to iterate all of
>   those other refs.
>
> That has _way_ less detail, but makes the point succinctly.

I don't disagree. Ultimately, it is a matter of maintainer preference, and I'm happy to follow (and instruct the AI to follow) the preferences described in this thread.

Show 8 quoted lines
>
> I dunno. I am not trying to pick apart your commit in particular, but am
> more interested in the broader use of AI commit messages going forward.
> This kind of verbosity is quite common in the output (from my limited
> experience), and I think creates more work for reviewers. Should we be
> expecting contributors to make things more concise before submitting
> (either manually or through prompting)? Or do people even agree that the
> shorter version is preferable? I could be the only one.

The AI does what you tell it; in this case I was telling it to follow the precedent in the repo and to ensure its claims are always cited. I'll tune it for succinct output going forward.

Show 93 quoted lines
>
> I have a few other comments on the patch itself below.
>
> > diff --git a/builtin/describe.c b/builtin/describe.c
> > index 1c47d7c0b7..3532c8ff22 100644
> > --- a/builtin/describe.c
> > +++ b/builtin/describe.c
> > @@ -740,6 +740,9 @@ int cmd_describe(int argc,
> >               return ret;
> >       }
> >
> > +     if (!all)
> > +             for_each_ref_opts.prefix = "refs/tags/";
> > +
> >       hashmap_init(&names, commit_name_neq, NULL, 0);
> >       refs_for_each_ref_ext(get_main_ref_store(the_repository),
> >                             get_name, NULL, &for_each_ref_opts);
>
> The code change looks fine. It creates a bit of a subtle dependency
> between what's happening here, and the filtering inside get_name(). But
> I think that's OK for the scope of a single command. It _might_ be
> possible to simplify the top of get_name(), since we'd no longer see
> non-tag refs in the input. But it also may not, since we have to strip
> out the prefix anyway. It can certainly come on top as a cleanup later
> if we want.
>
> > diff --git a/t/perf/p6100-describe.sh b/t/perf/p6100-describe.sh
>
> It is a little curious that we add a perf test here, but the commit
> message does not even show it off. ;)
>
> I ran it myself here and had trouble showing improvement, simply because
> it is already quite fast! I guess that's because I'm on Linux, where
> warm-cache filesystem operations are pretty fast. Bumping $ref_count by
> a factor of 10 made the "before" case 30ms, and after is still sub-1ms.
>
> > +test_expect_success 'set up many unrelated refs' '
> > +     ref_count=10000 &&
> > +     git tag -m tip tip HEAD &&
> > +     for i in $(test_seq $ref_count)
> > +     do
> > +             printf "create refs/heads/describe-perf/%05d HEAD\n" $i ||
> > +             return 1
> > +     done >instructions &&
> > +     git update-ref --stdin <instructions
> > +'
>
> A few things come to mind on reading this.
>
> I have mixed feelings on sticking synthetic constructions in the t/perf
> suite. Part of the original point was that we'd run it against real
> repos to see how they perform. But that implies that people running it
> have some clue about which tests may be interesting on which repos,
> which is hopeful at best. So we've turned to this kind of synthetic
> construction at times (and this is certainly not the first). It's
> probably a reasonable tactic here.
>
> I suspect the resulting state is not all that realistic, though. If you
> have 10,000 refs, you probably didn't make them all at once. And so in
> practice the majority of them would be packed. Sticking "git pack-refs
> --all" at the end might give more realistic numbers.
>
> Bumping to a larger number of refs shows the effect more clearly, but at
> the cost of making the setup take a long time (since we have to take a
> lockfile on each ref!). We could sneak around it by generating a
> packed-refs file directly, but now the test really would be
> backend-specific. Probably better not to go there.
>
> And finally, the loop can be written a bit more succinctly these days
> as:
>
> diff --git a/t/perf/p6100-describe.sh b/t/perf/p6100-describe.sh
> index ed9f1abe18..b365dc67ee 100755
> --- a/t/perf/p6100-describe.sh
> +++ b/t/perf/p6100-describe.sh
> @@ -30,12 +30,8 @@ test_perf 'describe HEAD with one tag' '
>  test_expect_success 'set up many unrelated refs' '
>         ref_count=10000 &&
>         git tag -m tip tip HEAD &&
> -       for i in $(test_seq $ref_count)
> -       do
> -               printf "create refs/heads/describe-perf/%05d HEAD\n" $i ||
> -               return 1
> -       done >instructions &&
> -       git update-ref --stdin <instructions
> +       test_seq -f "create refs/heads/describe-perf/%05d HEAD" $ref_count |
> +       git update-ref --stdin
>  '
>
>  test_perf 'describe exact tag with many unrelated refs' '
>
>
> Probably not worth re-rolling on its own, though.

The suggested changes seem reasonable to me. Certainly I am happy to make them, and re-rolls are cheap. Do let me know explicitly if you'd like that done.

>
> -Peff
Thanks for your time! I really appreciate it.
Previous: Jeff KingNext: Junio C Hamano
Message 6 of 12 in “describe: limit default ref iteration to tags”
  1. describe: limit default ref iteration to tagsTamir Duberstein, Jun 9, 2026
  2. Jeff KingJun 9, 2026
  3. Junio C HamanoJun 9, 2026
  4. D. Ben KnobleJun 9, 2026
  5. Jeff KingJun 11, 2026
  6. Tamir DubersteinJun 9, 2026
  7. Junio C HamanoJun 10, 2026
  8. Patrick SteinhardtJun 10, 2026
  9. Junio C HamanoJun 10, 2026
  10. Jeff KingJun 11, 2026
  11. describe: limit default ref iteration to tagsTamir Duberstein, Jun 10, 2026
  12. Jeff KingJun 11, 2026

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.