Patrick Steinhardt <ps@pks.im> writes:

I think you're right it's t5551/t5559. The runs Junio linked:

  osx-clang     cancelled  360min
  osx-gcc       cancelled  360min
  osx-reftable  success     35min
  osx-meson     success     61min

All four run the same t5551/t5559 under EXPENSIVE. The two that finished differ in just two ways, which look like the levers: osx-reftable generates the 100k-ref advertisement in ~24ms vs ~1.2s for loose refs on macOS (so much less time mid-response), and osx-meson runs tests at nproc while the prove jobs hardcode --jobs=10 on a 3-core runner (over recent master/next the prove jobs hang ~40%, meson ~10%).

When it is wedged the whole chain sits at 0% CPU. upload-pack is blocked in write() on the ls-refs advertisement, curl blocked in select(). So it looks like an HTTP/2 flow-control stall on the response side. The same stall resets itself after ~60-85s on my Linux box and on a bare-metal Mac, but not on the GitHub runner; I haven't pinned down why yet.

On the chance those two levers are the fix, a branch off master:

  https://github.com/mmontalbo/git/tree/mm/macos-ci-hang-fix

  - pack the refs in t5551's enormous-ref-negotiation test (doesn't
    change what it checks on the wire, just avoids re-reading 100k loose
    files to advertise them, like reftable already does)
  - use the core count for $JOBS on the GitHub macOS path, matching the
    GitLab branch in the same ci/lib.sh and what meson does

I ran the two macOS jobs under EXPENSIVE about eight times with these and they all finished in ~30-44min instead of hanging. Happy to send out a patch if it's helpful.

On Sat, Jun 20, 2026 at 08:33:13AM -0700, Michael Montalbo wrote:

If the problem is a racy deadlock, there is a reasonable chance that some jobs may simply be lucky. Even if things like packing refs help, I suspect the problem may still be lurking. Maybe I'm just a pessimist, though. ;)

We had some HTTP/2 stalls/deadlocks in the past, and they were dependent on libcurl and apache (actually h2_mod) versions. IIRC some of the non-TLS code paths for HTTP/2 were not well tested, which led to 8f2146dbf1 (t5559: make SSL/TLS the default, 2023-02-23). Of course after that commit those cleartext code paths should not be a problem, so that is probably not exactly the issue now.

But it might be worth checking the versions you're running locally versus what's in the GitHub runner.

-Peff

On Sun, Jun 21, 2026 at 05:34:07PM -0400, Jeff King wrote:

I had the same thought.

I didn't observe any similar hangs in GitLab's CI systems, so I wonder whether this is because of different versions of curl. And indeed we use different versions:

  - On GitHub we use 8.6.0.

  - On GitLab we use 8.7.1.

Now this of course doesn't mean that updating the curl version is the fix to this whole issue, as there's a ton of other factors that could play a role in whether or not the test hangs. So while we could just upgrade parts of the stack and cross our fingers, but that feels rather unsatisfactory. Still, one place to start could be to update our build images to macOS 15.

But the big question to me is whether the hang is because of a bug in Git with how we drive curl, a bug in curl itself, or a bug in Apache.

Patrick

Jeff King <peff@peff.net> writes:

I share the pessimism X-<.

True.

On Mon, Jun 22, 2026 at 06:42:24AM +0200, Patrick Steinhardt wrote:

[snip]

I noticed that a osx-clang job failed today in t5551 [1]. This time it didn't hang, but produced an actual error:

    2026-06-22T09:25:45.1984230Z ++ git -C too-many-refs fetch -q --tags
    2026-06-22T09:25:45.1984420Z error: RPC failed; curl 18 transfer closed with outstanding read data remaining
    2026-06-22T09:25:45.1984520Z fatal: expected flush after ref listing
    2026-06-22T09:25:45.1984610Z error: last command exited with $?=128
    2026-06-22T09:25:45.1984660Z ++ rm -f tags
    2026-06-22T09:25:45.1984710Z ++ :
    2026-06-22T09:25:45.1984830Z not ok 35 - http can handle enormous ref negotiation

There was a second test failing similarly.

Patrick

[1]: https://github.com/git/git/actions/runs/27940620478/job/82672854726

On Mon, Jun 22, 2026 at 11:48:01AM +0200, Patrick Steinhardt wrote:

Oh, and Linux is also failing in the same test suite [1], even though the job logs are truncated, so it's hard to say whether it's the same failure or not.

There certainly seems to be a deeper issue here. We could of course just disable the test again, but by now I do wonder whether this would paper over an actual bug.

Patrick

[1]: https://github.com/git/git/actions/runs/27940620478/job/82672854864

On Mon, Jun 22, 2026 at 11:55:31AM +0200, Patrick Steinhardt wrote:

Sorry for the repeated spam.

I think the issue is rather simple: we're hitting timeouts in Apache. If you apply the following diff:

Patrick Steinhardt <ps@pks.im> writes:

I think that is the trigger for issues we've been seeing. I spent some time investigating the Apache side over the last week and maybe found a mod_http2 bug, which I filed upstream with a potential fix:

  bug:  https://bz.apache.org/bugzilla/show_bug.cgi?id=70131
  fix:  https://github.com/mmontalbo/httpd/pull/2

To Patrick's earlier question of whether this is a Git, curl, or Apache bug: as best I can tell it's Apache. I could reproduce it with no Git involved at all (just Apache and a small CGI that goes quiet past the Timeout), and across several curl versions (8.6.0, which is what the GitHub runners use, up to 8.20.0), so I don't think bumping curl would help. It also seems to wear two faces from the same trigger: over HTTP/1.1 Apache closes the connection and curl bails with the "transfer closed" error (which looks like what you hit with Timeout=1, and the recent failures on both macOS and Linux), and over HTTP/2 it does not reliably reset the stream, so the client just waits, which is the six-hour macOS hang. I share the pessimism from earlier in the thread, though: I think the real fix is upstream in Apache, and anything we do on our side mostly just bounds the symptom in the meantime.

Given there could be a potential reliability issue with an upstream dependency like Apache, I was considering what mitigation strategies might help:

  - Enforce some kind of lower bound speed limit and a client-side
    timeout so runs that wedge fail fast (and loudly) instead of
    hanging.

  - Potentially provide some affordance for retrying flaky tests
    that might fail due to upstream dependencies. Git already has
    some HTTP retry support (http.maxRetries and friends, added
    recently), but as far as I can tell it only triggers on HTTP 429
    rate limiting, so it would not catch a stall like this on its
    own. A test-level retry is not something I like that much, since
    it might encourage papering over flakiness that should be
    resolved, but it was a consideration vs requiring a fresh CI run
    to resolve the flake.

  - Make slow tests faster by optimizing the test itself and/or
    the test runner configuration (e.g., job number matching
    cores) so wedges become less likely.

For the first one, I think Git already provides some affordances. There is a stall-based timeout that just ships disabled: as I understand it http.lowSpeedLimit sets a bytes/sec floor and http.lowSpeedTime how long a transfer can sit below it before curl gives up, so it would catch a wedged connection without punishing one that is just slow. Enabling it for the http tests might look something like:

    diff --git a/t/lib-httpd.sh b/t/lib-httpd.sh
    @@ GIT_TRACE=$GIT_TRACE; export GIT_TRACE
    +# Abort a transfer that makes essentially no progress for a while,
    +# so a wedged connection fails in seconds instead of hanging to the
    +# job cap. Tiny limit, generous window, so it only trips on a true
    +# stall; override either var, or set the limit to 0, to disable.
    +GIT_HTTP_LOW_SPEED_LIMIT=${GIT_HTTP_LOW_SPEED_LIMIT-1}
    +GIT_HTTP_LOW_SPEED_TIME=${GIT_HTTP_LOW_SPEED_TIME-60}
    +export GIT_HTTP_LOW_SPEED_LIMIT GIT_HTTP_LOW_SPEED_TIME

I went conservative on the values on purpose: a floor of 1 byte/sec should only really fire on a true zero-progress stall, not on something that is just crawling on a slow runner, and the 60s window is generous for the same reason. When I tried it locally against a stall-proxy it did turn an otherwise indefinite hang into a bounded abort (a tighter limit/window brings that down to single-digit seconds). It probably does not need to be suite-wide either; it could be scoped per-command with git -c, which the http tests already lean on for this kind of thing (t5551 passes http.postbuffer and http.extraheader that way), if a narrower blast radius feels safer.

I only dug into the first option in any depth, since I wanted to sanity-check the direction before writing patches. Does turning on a stall timeout for the http tests seem reasonable? Are there other strategies that we should implement?

On Thu, Jun 25, 2026 at 08:27:35PM -0700, Michael Montalbo wrote:

Thanks both of you for digging into this. I'm not familiar enough with Apache's code to pass confident judgement, but your findings certainly convinced me that this is just an apache bug.

Depending on how widespread the Apache bug is, another option might just be: do nothing and wait for it to get fixed.

Trying to make the wedged state fail fast and loudly is mostly just punting on the problem. We'd still see spurious failures. We've so far resisted the urge to do any automatic flaky-test retries, preferring instead to just try to root out the flakes. I'm a little hesitant to start now, because I think our strategy has mostly been good so far, and I've seen some horrible counter-examples where flakes and retries become a routine drag on development (and I'm afraid that accommodating flakes might make them more common).

It sounds like the bad state is triggered when Apache hits a timeout, and we hit that timeout because the system is slow or busy. We could try to make things less slow, but would it work equally well to increase that timeout?

-Peff

On Fri, Jun 26, 2026 at 01:16:57AM -0400, Jeff King wrote:

The bug manifests both with HTTP/1.1 and HTTP/2 though, so this wouldn't fully fix the flakes we see, right?

I agree. I'm not a fan of retry logic, as every flaky test may mask an actual bug that we haven't fully investigated yet.

I was also wondering whether we can maybe work around the issue by increasing the Apache timeout value. That sounds like an easy potential solution to try, and from all we've discovered so far it doesn't feel like this is something we can address on the Git side.

Patrick

Patrick Steinhardt <ps@pks.im> writes:

Can't agree more.

Thanks, all, for looking into this.

On Fri, Jun 26, 2026 at 3:50 AM Patrick Steinhardt <ps@pks.im> wrote:

Yes you are right. The linked fix would just prevent the hanging after timeout for HTTP/2 tests, but still leaves HTTP/1.1 fakes.

I think Peff and Patrick's suggestion to just increase the Apache timeout makes sense. I ran some experiments using a really long timeout with an artificially slowed down CI runner and all the jobs made progress (if slowly) without stalling, and eventually completed successfully:

https://github.com/mmontalbo/git/actions/runs/28267019651

I haven't spent a lot of time trying to figure out what the right timeout value should be. An hour definitely seems like overkill, with something on the order of 5-10 minutes seeming more reasonable, but I don't have a principled number.

On Fri, Jun 26, 2026 at 12:50:17PM +0200, Patrick Steinhardt wrote:

If I understand the situation correctly, there are really two problems.

First, the EXPENSIVE tests in t5551 sometimes trigger a timeout with Apache's stock settings. This presumably became a problem recently due to 7a094d68a2 (ci: run expensive tests on push builds to integration branches, 2026-05-08). The same problem exists in t5559, which just wraps t5551 but tells us to use http2.

This timeout will cause test failures in t5551, because we aren't able to complete a request we expected to. Obviously bad and annoying.

The second problem is that when Apache hits the timeout in HTTP/2 mode, it hangs forever. And then the CI job hangs for 6 hours until it's killed, which is an even more annoying failure.

So the root cause is the same (a timeout), but the effect depends on HTTP/1.1 vs HTTP/2. I was able to reproduce both cases on my local Debian unstable system by dropping the timeout as you suggested. Running t5551 with GIT_TEST_LONG yields a failure, whereas running t5559 yields a hang.

We can mitigate both cases by bumping the timeout value, since it's addressing the root cause.

There's an open question of whether this is just papering over a problem that real users might experience, and whether Git should be doing more to keep the connection alive. I think it's probably OK to ignore this in practice. This is an intentionally large request being served by a very underpowered platform. The default apache timeout is 60s. If a real-world server is seeing ls-refs requests take that long then they probably need to reconsider some other decisions, from ref packing to better hardware to dropping some users. ;) I don't think trying to insert keepalives at the Git layer here is worth the trouble.

To give a sense of the time options, here are a few timings from my local machine, timing "git upload-pack . </dev/null >/dev/null" in t5551's big repo.git (that's a v0 advertisement, but it should be roughly the same work as the v2 ls-refs).

  cold-cache, refs not packed:
  real	0m9.973s
  user	0m0.354s
  sys	0m1.364s

  warm cache, refs not packed:
  real	0m0.410s
  user	0m0.153s
  sys	0m0.257s

  cold-cache, refs packed:
  real	0m0.149s
  user	0m0.086s
  sys	0m0.035s

  warm cache, refs packed:
  real	0m0.069s
  user	0m0.054s
  sys	0m0.016s

So 10s is pretty abysmal (and on an SSD, no less). I would expect the cache to be warm (we just wrote these refs!) but I could also believe that CI systems are under heavy I/O and memory pressure, so we sometimes end up crossing the 60s mark.

So bumping Apache's timeout to 600s or something would probably be a fine mitigation. That's still not _solving_ the problem, but presumably an order of magnitude is enough for it to never come up in practice.

Michael suggested packing the refs as a mitigation. I was lukewarm on that in my previous email, because it wasn't clear to me how close we were on the timeout budget, and if it would just make the race less frequent (rather than never happen). But seeing those cold-cache numbers makes me think it might be worth doing just on principle to make the tests more efficient, and any timeout mitigation is a bonus.

Of course the pack-refs process (and the initial ref writes) will still have to touch all of those loose files, so those will still be slow. But they're not on a timeout, and I suspect we read the result many more times than we write/pack (the test failures we are seeing are not in the expensive tests, but just "normal" tests that are stuck with the gigantic ref state).

-Peff

On Fri, Jun 26, 2026 at 04:26:28PM -0700, Michael Montalbo wrote:

Here are some patches to keep things moving along. I arbitrarily picked 10 minutes, because multiplying the 1-minute default by 10 felt right. ;)

The first one just bumps the timeout and should make our problems go away. The other two are optimizations, but I'm on the fence on whether the final patch is worth it.

Thanks again for all of the digging.

  [1/3]: t/lib-httpd: bump apache timeout
  [2/3]: t5551: put many-tags case into its own repo
  [3/3]: t5551: pack refs after creating many tags

 t/lib-httpd/apache.conf     |  1 +
 t/t5551-http-fetch-smart.sh | 10 ++++++----
 2 files changed, 7 insertions(+), 4 deletions(-)

-Peff

Since enabling more tests with 7a094d68a2 (ci: run expensive tests on push builds to integration branches, 2026-05-08), we sometimes see test failures or timeouts in GitHub CI. The culprit seems to be the "enormous ref negotiation" test in t5551, which creates ~100k tag refs in our http server-side repo.

Iterating through the loose refs of this repo to generate a ref advertisement can take a long time, especially on a platform with slow I/O. On my otherwise unloaded local machine, a cold cache ref advertisement takes ~10s. On a busy CI machine running tests in parallel, it can presumably top 60s, which runs afoul of Apache's default CGI timeout.

The result in t5551 is a test failure, where Apache simply hangs up the connection and the client reports an error. But worse, t5559 runs the same test with HTTP/2, and a bug in Apache causes the connection to hang indefinitely! We eventually see this as a CI timeout after 6 hours.

Let's bump Apache's timeout to something much larger: 600 seconds. This doesn't eliminate the possibility of a timeout, but it makes it much less likely. It should eliminate both the test failures and the CI timeouts in practice, and it protects us from running into similar problems with other tests in the future.

There are two counter-arguments to consider.

One, could/should we just make the test faster? Probably yes. The biggest mistake here is having such an absurd number of unpacked refs on a system which is bottle-necked on I/O. But I think it's worth bumping the timeout so that we can fix this (and possibly other) correctness issues, and then consider performance separately (which we'll do in subsequent patches).

And two, is this just papering over a problem that users might see in the real world? We could teach Git to handle this case more gracefully with optimizations or keep-alives. But I think it's really an artificial situation. You need a combination of this silly number of loose refs, plus a very heavily loaded system. If you were trying to run a real server and it took more than 60s to generate the ref advertisement, I don't think the timeout is your biggest problem. Your crappy service is, and you should adjust your resources to match your load. I.e., it is probably reasonable for Git to assume that advertisements happen fast-ish and don't need protocol-level keepalives.

Though the patch here is small, tons of work went into analyzing the problem. Many thanks to the contributors credited below.

Helped-by: Michael Montalbo <mmontalbo@gmail.com>
Helped-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Jeff King <peff@peff.net>
---
I didn't reference Michael's bugzilla report directly, because you can't
read it without a login. :(

Maybe it's worth doing anyway?

 t/lib-httpd/apache.conf | 1 +
 1 file changed, 1 insertion(+)

Most of the t5551 http fetch tests use a handful of refs. But there are a few test cases which check our handling of large numbers of refs. These tests use the same server-side repo, so all subsequent tests end up having to consider those extra refs, too.

The result is that the test script is a bit slower than it needs to be. In a normal run, moving the "2,000 tags" test into its own repo drops my runtime for the whole script from ~2.7s to ~1.9s.

This is a modest gain, but when we add the "--long" flag it gets much bigger. There we trigger a test (marked with EXPENSIVE) that adds 100,000 tags, and the script runtime jumps to ~95s. But if we use the same "many tags" repo for that, our runtime drops to just ~37s.

This is a pretty easy win to drop the cost of the script. It may even be a larger gain on a heavily loaded system, since one of the main costs here is unpacked refs, which are heavy on system time and I/O costs.

It's possible we are reducing test coverage, since all of those other tests were inadvertently using large ref advertisements (and thus could have uncovered some unexpected interaction). But that seems somewhat unlikely; the tests targeted at the large number of refs are doing roughly similar things to the other tests.

Note that the real performance culprit is the 100k-tag --long test, not the 2k-tag one. So we could just let the 100k one use its own repo, and keep the 2k tags in the main repo. But since these two tests are somewhat interlinked, it's easier to just move them both (and it does provide a small gain even for the 2000-tag test). I also notice that the 2000-tag test is gated on the CMDLINE_LIMIT prereq, and without that the later EXPENSIVE test will fail (since we won't have a too-many-refs clone). Nobody seems to have noticed or complained after many years, and I left it alone for this patch.

Signed-off-by: Jeff King <peff@peff.net>
---
 t/t5551-http-fetch-smart.sh | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

We have two tests that create 2,000 and 100,000 tags respectively. After doing so, the resulting state can be a bit slow to work with when using the "files" ref backend, as each of those refs is in its own file.

This isn't a very realistic scenario, as we'd expect most of those refs to be packed. If they accrue over time along with objects, they'd get packed by maintenance/gc runs. And if you have a process that creates a ton of refs at once (like a big fast-import), the usual recommendation is to run maintenance afterwards.

So let's follow that recommendation and pack the refs ourselves. Unfortunately, this does not seem to produce an improvement to the run-time of the test script! That's because after producing this state, we perform only a few fetches of it. And packing the refs costs at least as much as serving a ref advertisement (both have to iterate the refs, but packing additionally must write .lock files as we pack).

My wall-clock time was slightly improved (but within the noise) with this patch, but my user and system CPU time were slightly worse! However, on a loaded system with I/O bottlenecks, it may be a net win. That's somewhat of a guess, though.

It would be nice if we had a way to generate all of these refs without writing so many individual files. But even if we taught the ref code to write large cases directly to the packed-refs file, we'd still need to take individual locks. The real solution is a backend like reftable, which shaves ~30% off of the test runtime.

Signed-off-by: Jeff King <peff@peff.net>
---
I'm iffy on whether this one is worth it.

If you apply just this patch without patch 2, then the run-time does improve quite a bit. The cost of packing is amortized by the improved performance for all of those subsequent tests (but after patch 2, they never even see the unpacked state).

Likewise, I suspect this would make our timeout problems go away even without patch 1.

So the whole series _could_ be reduced to just this one patch. But hopefully the reasoning given in the earlier patches makes sense, at which point this one is kind of superfluous.

 t/t5551-http-fetch-smart.sh | 1 +
 1 file changed, 1 insertion(+)

Jeff King <peff@peff.net> writes:

Testing a pathological set-up with too many loose refs may have extra value, as long as we are also testing the recommended set-up, so ideally we should have both ;-) but if we have to pick only one and drop the other, we probably should be testing the packed case.

I am ambivalent, too, about this change for the purpose of the "yeek, apache times out while enumerating refs" issue. But see above ;-)

Jeff King <peff@peff.net> writes:

While all the other repositories used in this tests are bare repositories, this new one is a non-bare repository.

It shouldn't make any difference, but since I noticed it...

On Sun, Jun 28, 2026 at 02:44:32PM -0700, Junio C Hamano wrote:

Ah, yeah. It should work either way, but it is slightly confusing for it to be non-bare. I'll wait to re-send (though if nothing else comes up, it may be simpler for you to just amend on your side).

-Peff

On Sun, Jun 28, 2026 at 04:07:10AM -0400, Jeff King wrote:

We kind of already have this with the `REF_TRANSACTION_FLAG_INITIAL` flag, but right now it is only used when performing a clone or when migrating references. Also, it requires an empty repository that has no references yet.

It raises the question whether we could also extend git-fast-import(1) to use it, as it would typically be run on an almost-empty repository. It's the "almost" that kills it though, as we already do have at least the HEAD reference. So it could be feasible, but it's not as trivial as just setting the flag and then we're magically faster.

And besides, in this particular test here we run git-fast-import(1) multiple times in the same repository, so it wouldn't help us.

We could of course extend all of this so that Git is able to write into the packed-refs directly, even with preexisting refs. But I agree with your sentiment: it doesn't feel worth it as the reftable backend fixes scenarios like this anyway.

Agreed. I'd just merge the first two patches and drop this one here.

Thanks!

Patrick

On Sun, Jun 28, 2026 at 03:57:16AM -0400, Jeff King wrote:

By the way, the only reason why we at GitLab haven't been feeling the pain is that we only enable GIT_TEST_LONG for GitHub. So I was wondering whether we want to have something like the below patch on top.

Patrick

Patrick Steinhardt <ps@pks.im> writes:

If we can afford the cycles, it would be good to have similarly larger coverage on two different platforms (compared to leaving one of them not doing as much as the other when we know it). On the other hand, if we cannot cover _everything_ in one platform, it may be a better use of the resources to have the other platform things that are not covered already. I see that among different pipeline sources, we are doing TEST_LONG for pull requests to any branch, and pushes only to "cast in stone" branches. If there are other branches that deserve to be tested with TEST_LONG upon other events that the existing GitHub Actions CI does not trigger, it may be good to have GitLab CI cover them, perhaps?

Jeff King <peff@peff.net> writes:

OK. It seems both Patrick and you are in favor of using only [1/3] & [2/3] but dropping [3/3]? If that is the concensus I can just tweak this one and apply before 2.55 final.

Thanks.

On Mon, Jun 29, 2026 at 07:39:59AM -0700, Junio C Hamano wrote:

I'm a bit hesitant to do such a split, mostly because the canonical source of truth that the project typically uses is GitHub's CI. So I want us at GitLab to be able to catch the same issues that GitHub would flag. And if GitLab's CI stopped detecting everything that GitHub does, then the result would likely be that we often create merge requests on both platforms, which would only result in more wasted resources.

Patrick

Patrick Steinhardt <ps@pks.im> writes:

I didn't suggest splitting them into two circles that overlap but each with area only it covers, though. GitLab's coverage can be superset to GitHub's and that would satify what I suggested.

FWIW, I do not consider GitHub's CI "the canonical source" at all. It is a very handy service to use to check how well we are doing, but from time to time it has its own hiccups ;-).

What can we do to make the visibility of GitLab's CI more prominent?

I know where the CI jobs that are triggered when I push out the integration branches are found at GitHub's website[*], but I do not think I know the corresponding one at GitLab, for example, and I think that is a shame.

[Footnote]
 *1* I just made https://tinyurl.com/github-gitci that points at
     https://github.com/git/git/actions/workflows/main.yml?query=event%3Apush+actor%3Agitster

On Mon, Jun 29, 2026 at 07:57:21AM +0200, Patrick Steinhardt wrote:

Yup. In the past I've pondered exposing this via update-ref, but I think it's too weird and/or dangerous to do so. Especially because you are still stuck creating all of the .lock files, so the performance is not even that much better (though it does save you doing so _twice_ when you then pack the refs).

So the performance option you really want is "YOLO, just write some packed-refs without locking". But that is not something I think we want to expose to users. ;)

We could do it ad-hoc within this test like so:

On Mon, Jun 29, 2026 at 07:42:31AM -0700, Junio C Hamano wrote:

Yep, that sounds great. Looks like it already happened. :)

-Peff

On Mon, Jun 29, 2026 at 09:19:11AM -0700, Junio C Hamano wrote:

Fair.

Well, GitLab of course has its own share of hiccups, like for example the Chocolatey issues we've been facing.

The pipelines of the official mirror can be found at [1]. We might for example add something like the below patch to our README.md to make it more discoverable.

Patrick

[1]: https://gitlab.com/git-scm/git/-/pipelines

On Mon, Jun 29, 2026 at 04:35:27PM -0400, Jeff King wrote:

Yeah, this rings a bell. If I remember correctly, this is because we call `refs_resolve_ref_unsafe()` to verify whether the target already exists. And as that function is generic, it wasn't easy to optimize it by reusing the already-loaded reftable stack.

Yeah, that would certainly be an option to explore.

True, it's going to be much better compared to the "files" backend. But that isn't enough reason to not optimize it even further -- doubly so if it actually takes 85% of the time. Sounds like a low-hanging fruit to me that can result in a significant speedup.

I'll probably not get to it anytime soon, but I'll create an issue to keep track of it.

Patrick

Patrick Steinhardt <ps@pks.im> writes:

Oh, nice. We of course do not want to be heavily involved in advertising offerings by commercial entities but I think these two sites deserve one line each for their continued service to the community ;-)

On Tue, Jun 30, 2026 at 11:05:22AM +0200, Patrick Steinhardt wrote:

I think it's similar to that. I dug down and got some actual performance numbers below for eliminating the fstat() calls, with the conclusion that it's probably not worth spending too much time digging into it. But details below for completeness.

I was wrong to say fstat() before, it's a regular stat(). The interesting backtrace is:

  #0  __GI___stat64 (file=0x555555ab0bf0 "/home/peff/tmp/.git/reftable/tables.list", buf=0x7fffffffd630)
      at ../sysdeps/unix/sysv/linux/stat64.c:29
  #1  0x0000555555832ad6 in stack_uptodate (st=0x555555ab0ae0) at reftable/stack.c:575
  #2  0x0000555555832c5c in reftable_stack_reload (st=0x555555ab0ae0) at reftable/stack.c:625
  #3  0x000055555581f690 in backend_for (out=0x7fffffffd798, store=0x555555ab0900,
      refname=0x555555aae178 "refs/tags/foo-1", rewritten_ref=0x7fffffffd780, reload=1) at refs/reftable-backend.c:283
  #4  0x0000555555824957 in reftable_be_reflog_exists (ref_store=0x555555ab0900,
      refname=0x555555aae178 "refs/tags/foo-1") at refs/reftable-backend.c:2266
  #5  0x000055555581393a in refs_reflog_exists (refs=0x555555ab0900, refname=0x555555aae178 "refs/tags/foo-1")
      at refs.c:2995
  #6  0x000055555581f730 in should_write_log (refs=0x555555ab0900, refname=0x555555aae178 "refs/tags/foo-1")
      at refs/reftable-backend.c:301
  #7  0x00005555558224df in write_transaction_table (writer=0x5555569abe30, cb_data=0x5555569ab820)
      at refs/reftable-backend.c:1511
  #8  0x000055555583354d in reftable_addition_add (add=0x5555569ab050,
      write_table=0x5555558220f0 <write_transaction_table>, arg=0x5555569ab820) at reftable/stack.c:902
  #9  0x0000555555822b2b in reftable_be_transaction_finish (ref_store=0x555555ab0900, transaction=0x555555ab0c80,
      err=0x7fffffffdbc0) at refs/reftable-backend.c:1633
  #10 0x0000555555813161 in ref_transaction_commit (transaction=0x555555ab0c80, err=0x7fffffffdbc0) at refs.c:2769
  #11 0x00005555556956c9 in update_refs_stdin (flags=0) at builtin/update-ref.c:789

So we are asking about reflogs for each ref under the "only reflog if a log already exists" rule. Which means we can easily disable it by setting core.logallrefupdates to "always", giving us a way to measure the impact. So we can try:

  git init --ref-format=reftable
  blob=$(echo foo | git hash-object -w --stdin)
  seq -f "create refs/tags/foo-%g $blob" 50000 >input
  cp -a .git/reftable reftable.orig
  hyperfine \
  	-p 'rm -rf .git/reftable; cp -a reftable.orig .git/reftable' \
  	-L config true,always \
  	'git -c core.logallrefupdates={config} update-ref --stdin <input'

which yields:

  Benchmark 1: git -c core.logallrefupdates=true update-ref --stdin <input
    Time (mean ± σ):     128.8 ms ±   1.5 ms    [User: 97.1 ms, System: 31.7 ms]
    Range (min … max):   126.4 ms … 131.3 ms    23 runs
  
  Benchmark 2: git -c core.logallrefupdates=always update-ref --stdin <input
    Time (mean ± σ):     195.2 ms ±   1.7 ms    [User: 182.1 ms, System: 13.0 ms]
    Range (min … max):   191.9 ms … 197.6 ms    15 runs

So we saved all of those stat() calls, but writing the actual reflog entries adds much more cost!

Sadly there is no "never" option for core.logallrefupdates. I hacked one in and the result took ~96ms to run. So that tells us the cost of the stat() checks: around 25% of the runtime.

Which is a big-ish percentage, but a small absolute number. It's 0.64us per ref. Perhaps not worrying about too much.

The other interesting thing I noticed is that we seem to write() entries individually with no buffering. Probably we could get an easy speedup for big cases like this by using stdio in the fd_writer abstraction. We're again pinching pennies to some degree, but I expect we may be able to drop the 120ms case down to 60ms or so (just guessing based on the extra reflog writes adding ~70ms).

There was one other oddity I didn't quite resolve. You may notice the gross reftable.orig stuff in hyperfine. I originally wrote this as:

    git for-each-ref --format="delete %(refname)" | git update-ref --stdin

but for some reason that causes the subsequent update-ref to loop infinitely on merged_iter_next_entry(). It does so reliably, but I can't reproduce it outside of hyperfine. Super weird, and I'm sure I'm missing something obvious.

-Peff

On Tue, Jun 30, 2026 at 07:47:02PM -0400, Jeff King wrote:

Ah, maybe not infinite, but probably quadratic. The key is that you have to delete a lot of refs and then try to insert them again. So with this script:

  nr=$1; shift
  rm -rf .git
  
  git init --ref-format=reftable
  blob=$(echo foo | git hash-object -w --stdin)
  seq -f "create refs/tags/foo-%g $blob" $nr >input
  git update-ref --stdin <input
  git for-each-ref --format="delete %(refname)" | git update-ref --stdin
  time git update-ref --stdin <input

I get results like this:

  nr   | runtime
  ------------
  1000 | 0.125s
  2000 | 0.454s
  4000 | 1.811s
  8000 | 7.091s

So for every doubling of the input size, the runtime quadruples. I guess it is iterating through some deleted tombstone entries, but I'm not sure why.

That's probably a more interesting and productive performance problem to work on than micro-optimizing out the last few microseconds of writing. :)

-Peff

On Tue, Jun 30, 2026 at 07:58:50PM -0400, Jeff King wrote:

This is a known issue, I think [1].

The problem here is the tombstoning: when you delete all references, chances are that they are not truly gone but that every reference is just tombstoned. The problem with this is that reading refs may now take signifciantly more time as we cannot just say "this stack is empty". Instead, we need to figure out that it is empty by processing all the tombstones, and that takes a lot of time.

I remember that I did some digging back then and improved the status quo quite significantly by optimizing `refs_verify_refname_available()`. I'm sure there are more opportunities for optimization here though -- I have a feeling that we for example exhaust the merged iterator until its end when searching for a specific refname, where we could easily abort once the observed tombstone name sorts lexicographically after the needle.

But eventually I decided to not care too much about this edge case, as it seems very specific to this artificial benchmark scenario. Which of course doesn't mean that it's not worth doing, I just had bigger fish to fry and didn't get around to it yet.

Patrick

[1]: <Z602dzQggtDdcgCX@tapette.crustytoothpaste.net>

On Tue, Jun 30, 2026 at 12:21:30PM -0700, Junio C Hamano wrote:

Okay, I'll send a patch then. Thanks!

Patrick

On Wed, Jul 01, 2026 at 08:17:45AM +0200, Patrick Steinhardt wrote:

Yeah, it is (mostly) the same problem. About half the time is spent in refs_verify_refnames_available().

The other half is in reftable_be_transaction_prepare(). Looks like it makes individual calls to prepare_single_update(), which reads each ref. And those reads are expensive because of all of the tombstones. It might be possible to do an iterator merge or similar between the sorted list of transaction refs and the reftable contents.

Yeah, that's fair. I dug a bit further in case there was anything useful to write up, but I don't have much to add beyond what's here and in the thread you linked. We can let it live on in the archive for now.

-Peff

On Wed, 1 Jul 2026 at 10:00, Jeff King <peff@peff.net> wrote:

Hi, sorry for jumping in -- I found this interesting and started poking at the code. I think both halves may share the same root cause.

The merged iterator's suppress_deletions flag filters out tombstones internally, which means higher-level code with prefix or refname bounds never gets a chance to stop iteration early. By letting tombstones pass through and filtering them one layer up in the reftable backend, the existing bounds checks can kick in before we scan through all the tombstones.

So instead of doing full scans inside merged_iter_next_void() we can just delegate to merged_iter_next_entry() and instead add a loop to reftable_be_reflog_exists() that skips tombstones (but is amortized O(1)).

Now multiple call sites would need to add something like this to compensate for returning tombstones:

    if (reftable_log_record_is_deletion(&iter->log))
        continue;

but it may be worth it if it reduces cost when there are many refs.

The key spot is reftable_ref_iterator_advance(), where the deletion skip goes right after the existing prefix check -- so a tombstone past the prefix stops iteration immediately instead of being silently consumed. The same idea applies to reftable_backend_read_ref() and the log iteration paths.

I have a local branch with this attempted fix. Rerunning the benchmark:

  Before:
    nr=1000  0.306s
    nr=2000  0.945s
    nr=4000  3.816s
    nr=8000  14.93s

  After:
    nr=1000   0.020s
    nr=2000   0.044s
    nr=4000   0.071s
    nr=8000   0.145s
    nr=16000  0.258s
    nr=32000  0.591s

I can send a proper patch if needed/wanted, but I might have missed something silly here.

Thanks, Kristofer

On Wed, Jul 01, 2026 at 11:04:45AM +0200, Kristofer Karlsson wrote:

Ah, interesting. This kind of hits int o the direction that I proposed of stopping iteration once we hit the first reference that has a lexicographically-larget refname. But I thought about having to solve it generically in the iterator somehow, and I wasn't really looking forward to that.

Your solutions works around that issue indeed. It of course doesn't solve reference iteration, which would still be expensive. But it does solve the issue where we just want to look up a single reference, which is what Peff noticed to be slow.

I think it certainly might be.

Yup.

Nice gains. I certainly think it would make sense to polish this a bit and then cast it into a patch.

Patrick

On Sun, Jun 28, 2026 at 1:00 AM Jeff King <peff@peff.net> wrote:

I also thought the report being behind a login was unfortunate. For the historical record, I ended up submitting a patch[1] to their public GitHub mirror that describes the issue in more detail.

[1] https://github.com/apache/httpd/pull/676

On Wed, 1 Jul 2026 at 12:07, Patrick Steinhardt <ps@pks.im> wrote:

I have a small draft here https://github.com/gitgitgadget/git/pull/2166 but I am honestly not sure if it's worth submitting as a patch - the change is somewhat small, but spread out, and I failed to properly reproduce the performance win in any realistic scenario (I had to disable compaction to see the improvement).

I would want to rely on your expertise to know if this change would be valuable to discuss as a patch at all.

Thanks, Kristofer

On Fri, Jul 03, 2026 at 02:09:45PM +0200, Kristofer Karlsson wrote:

An easy scenario where you don't have to disable compaction would be what Peff posted: you create X references and then delete all of them. That shouldn't result in compaction and directly hits the case that we care about.

If we can demonstrate a significant improvement in the above case then it would be worth it, I guess.

Patrick

On Mon, 6 Jul 2026 at 08:46, Patrick Steinhardt <ps@pks.im> wrote:

Right, thanks. The recreate-same-refs case works with compaction enabled and shows the expected improvement (~100x for 8000 refs).

I also found a worse case that feels more realistic: delete 8000 "old-*" refs, then create 8000 "new-*" refs. Since "new" is lexicographically after "old", every create scans all tombstones. That one goes from 27s to 0.09s after fixing it.

I will clean up the patch and submit it shortly.

Thanks, Kristofer