Re: [Internet]Re: [PATCH v2] send-pack.c: add config push.useBitmaps

From: kylezhao (赵柯宇) <kylezhao@tencent.com>
Date: Jun 16, 2022, 15:11 UTC
Message-ID: <be5afab25a5b44b3af797c0c8b75f5e6@tencent.com>
In-Reply-To: <220616.86fsk4ww69.gmgdl@evledraar.gmail.com>

Show 7 quoted lines

> And since this is linking to that old thread I started in 2019 I really
>think the commit message should be exploring and justifying why this
> might be slower in all cases, or at least a sufficient number of cases
> to flip the default.
> 
> if it's not are we throwing out a generally useful optimization (by
> default) due to some edge cases we've found?

I'm not sure if it's an edge case. However, if a git repo has this problem, it's "git push" will always be very slow instead of going fast and slow.

> E.g. have you tried to follow-up on this comment by Jeff King?
> https://lore.kernel.org/git/20190523113031.GA17448@sigill.intra.peff.net/

Not yet, I'll try it later.

> At the time I didn't, because as noted in a follow-up I'd lost my test
> case by the time I read that, but it seems you haven't, and have a
> current test case.

I tried to find a test case in open-source projects, and finally found one. $ git clone https://github.com/JetBrains/intellij-community.git --bare $ cd intellij-community.git $ git repack -adb $ GIT_TRACE=1 git push . master:test1

Then it would get stuck in "git pack-objects" for about 10s. (OS: Windows) After I deleted the bitmap, "git push" took less than 1s.

> I think a really good start would be to come up with some benchmarks for
> a few common cases, e.g. how do bitmaps do if we have a big repo but few
> refs, how about a lot of refs? Does their stale-ness make a difference
> or not (per previous discussions) etc. etc.

According to my observation, the problem always occur on git repos which have a lot of refs and objects. Also, "git push" takes much time on "find _objects(..) for haves_bitmap".

Hope the information above can help find the cause of the problem.

Thanks,
- Kyle