Bisect limited to Merge Commits

Hey,

are there any plans to add an option to skip all non-merge commits via bisecting?

git bisect start --merges-only

Imagine a "rebase feature branch" style of development. All features are developed on separate features branch which are rebased on master and immediately merged into the upstream master. Now think about lazy programmers who miss to check the build process at each commit - leading to many unbuildable commits. The only guaranteed buildable commit is often the merge commit.

Bisecting only merge commits will reduce the amount of skip commands (think about many unbuildable commits) and reduce the number of tests.

Any comments on this idea? What about extending bisect.c:do_find_bisection()

	hgn

Hagen Paul Pfeifer <hagen@jauu.net> writes:

I do not want to imagine such ;-) The only semi-sensible reason why people might want to always rebase on top of 'master' is to keep the history completely linear, so with such a workflow, there won't be a merge commit.

But I think "rebase" part of your description is a red herring. If a development goes by always doing a new feature in a side branch and then merge the branch only after it is done to the 'master' branch, then bisecting only the commits on first-parent chain would often be a quick first-pass to find the topic whose merge into the 'master' branch introduced a breakage. From there you can dig down to each individual commit on the side branch. And for that, it is immaterial that the side branch gets rebased on 'master' and forced to become a merge with "--no-ff", of the side branch was developed on older upstream but in a careless way full of "oops, previous one I broke the entire world and it does not even compile; this commit fixes it" commits.

So being able to stop at only commits on the first-parent chain is a valid and useful tool. "git bisect --first-parent" is one of the things that are sometimes asked for.

Am 27.04.2016 um 22:56 schrieb Junio C Hamano:

With origin pointing to git.git, I attempted this:

  git bisect start
  git rev-list --first-parent --boundary origin..origin/pu |
    sed -ne s/-//p | xargs git bisect good
  git bisect bad origin/pu

and it starts bisecting among the 50-something first-parent commits between origin and origin/pu.

-- Hannes

Johannes Sixt <j6t@kdbg.org> writes:

That is a cute idea but I wonder if it would work well in a history full of pointless no-ff merges.

Here is a sample topology (output from "git log --oneline --graph")

    *   84ef1bb pointless
    |\  
    | * 4ae9f68 third
    |/  
    *   aec8732 pointless
    |\  
    | * fd4ed0d second
    |/  
    * 696b5c1 initial

where the tip of 'master' was initial, a side branch built 'second' on it and then 'master' made a no-ff 'pointless' merge, another side branch built 'third' on it and then 'master' again made a no-ff 'pointless' merge.

    $ git rev-list --first-parent --boundary initial..pointless

would give us 'third' and 'second' as boundaries, but by marking 'third' as GOOD, we declare to 'bisect' machinery that it and all of its ancestors are GOOD. So upon seeing 'bisect bad HEAD', we would immediately see that HEAD is the first bad commit, wouldn't we?

Actually, "pointless no-ff merges" is a red herring. Any history with side branch that branched from the first-parent chain after the commit at the bottom of the bisection range (in this example, the side branch that built 'third' was forked off of the first 'pointless' commit on the first-parent chain of 'master', which was made after the bottom of the range 'initial') would have the same problem, I would imagine.

* Johannes Sixt | 2016-04-27 23:33:53 [+0200]:

Hey Junio, hey Hannes,

just for clarification: contributors rebase their work before pushing it on master. The integrator simple merges --no-ff the individual branches. Just a regular workflow, nothing special - except that many contributor commits will not build. ;(

The idea is just to skip the contributor commits during bisect and focus on the merge commits (the ones with more than one ancestors) because they are likely build and testable.

One possible approach is probably to sort out all non-merge commits before bisecting and bisect only on a this set of commits. The advantage is that the first bad commit is the merge commit introduced the regression. Mmmh, any comments?

hgn

On Wed, Apr 27, 2016 at 11:19 PM, Hagen Paul Pfeifer <hagen@jauu.net> wrote:

I suspect doing something akin to the idea of "bisect --first-parent" would work for this use case and be more flexible in general. Your idea is pretty much what i think bisect --first-parent would do, except that it would also work for non-merge commits which happen to be in the "mainline" history.

Thanks, Jake