Re: [PATCH v2 3/3] connected: implement connectivity check using bitmaps

From: Patrick Steinhardt <ps@pks.im>
Date: Jul 20, 2021, 14:26 UTC
Message-ID: <YPbdJFfykDo1diJK@ncase>
In-Reply-To: <YNvOJQDYRWpFa+4S@coredump.intra.peff.net>

On Tue, Jun 29, 2021 at 09:51:33PM -0400, Jeff King wrote:

Show 29 quoted lines

> On Mon, Jun 28, 2021 at 07:33:15AM +0200, Patrick Steinhardt wrote:
> 
> > As expected, performance doesn't change in cases where we do not have a
> > bitmap available given that the old code path still kicks in. In case we
> > do have bitmaps, this is kind of a mixed bag: while git-receive-pack(1)
> > is slower in a "normal" clone of linux.git, it is significantly faster
> > for a clone with lots of references. The slowness can potentially be
> > explained by the overhead of loading the bitmap. On the other hand, the
> > new code is faster as expected in repos which have lots of references
> > given that we do not have to mark all negative references anymore.
> 
> Hmm. We _do_ still have to mark those negative references now, though
> (the bitmap code still considers each as a reachability tip for the
> "have" side of the traversal). It's just that we may have to do less
> traversal on them, if they're mentioned by other bitmaps.
> 
> So in that sense I don't think your "a ref for every commit" cases are
> all that interesting. Any bitmap near the tip of history is going to
> include a bit for all those old commits, because our fake set of refs
> are all reachable. A much more interesting history is when you have a
> bunch of little unreachable spikes coming off the main history.
> 
> This is common if you have a lot of branches in the repo, but also if
> you maintain a lot of book-keeping refs (like the refs/pull/* we do at
> GitHub; I assume GitLab does something similar).
> 
> Here are some real-world numbers from one of the repos that gives us
> frequent problems with bitmaps. refs/pull/9937/head in this case is an
> unmerged PR with 8 commits on it.

Yeah, this kind of brings us back to the old topic of pathological performance combined with bitmaps. As I said in the cover letter, I haven't been particularly happy with the results of this version, but rather intended it as an RFC. Taylor's extension does look quite interesting, but ultimately I'm not sure whether we want to use bitmaps for connectivity checks. Your spiky-branches example neatly highlights that it cannot really work in the general case.

I wonder where that leaves us. I'm out of ideas on how to solve this in the general case for any push/connectivity check, so I guess that any alternative approach would instead make use of heuristics.

In the current context, I care mostly about the user-side context, which is interactive pushes. Without knowing the numbers, my bet is that the most frequent usecase here is the push of a single branch with only a bunch of commits. If the pushed commit is a descendant of any existing commit, then we can limit the connectivity check to `git rev-list --objects $newoid --not $oldoid` instead of `--not --all`. There's two issues:

    - "descendant of any existing commit" is again the same territory
      performance-wise as `--all`. So we can heuristically limit this
      either to the to-be-updated-target reference if it exists, or
      HEAD.

    - Calculating ancestry can be expensive if there's too many commits
      in between or if history is unrelated. We may limit this check to
      a small number like only checking the most recent 16 commits.

If these conditions hold, then we can do above optimized check, otherwise we fall back to the old check.

Do we actually gain anything by this? The following was executed with linux.git and stable tags. afeb391 is an empty commit on top of master.

    Benchmark #1: git rev-list afeb391 --not --all
      Time (mean ± σ):      64.1 ms ±   8.0 ms    [User: 52.8 ms, System: 11.1 ms]
      Range (min … max):    58.2 ms …  79.5 ms    37 runs

    Benchmark #2: git rev-list afeb391 --not master
      Time (mean ± σ):       1.6 ms ±   0.5 ms    [User: 1.0 ms, System: 1.0 ms]
      Range (min … max):     0.4 ms …   2.2 ms    1678 runs

Obviously not a real-world example, but it serves as a hint that it would help in some cases and potentially pay out quite well.

Patrick