git bisect should return 1 when the first bad commit is found

Hi

I'm using git bisect, and for complex reasons I can't use git bisect run to drive the bisection. I don't think git bisect bad/good is returning a different status code when the first bad commit is found and this would be very useful to stop the driver of the bisection.

What do you think?  Would such a change be acceptable?

Thank you!

Pedro.

Pedro Larroy <pedro.larroy.lists@gmail.com> writes:

Probably not.

A non-zero exit code would indicate that "git bisect" had trouble finding the next commit to check after being told 'good' (or 'bad'), which you wouldn't be able to tell from your "first bad one is found".

The 'first bad one' is not necessarily the HEAD when the bisect procedure finds it, so you'd have to read and parse "x is the first such commit" message anyway, no? And once you start parsing the message from 'git bisect', you should be able to tell between "please try this one, you have approximately 7 steps left" and "we found the first bad one, which is X", without relying on the exit status, no?

Thanks for your answer.

I was expecting the HEAD to point to the first bad commit.

In mercurial, the exit status tells you information about the bisection process: https://www.mercurial-scm.org/repo/hg/help/bisect

Sure one can parse stdout, it's just more tedious than just checking the return code and having the HEAD left to the original bad commit.

Pedro.

On Thu, Jun 13, 2019 at 11:22 AM Junio C Hamano <gitster@pobox.com> wrote:

On Sun, Jun 23, 2019 at 01:32:16PM -0700, Pedro Larroy wrote:

I think it might be nice for Git to write a well-known refname (like BISECT_RESULT or similar) so that you can refer to that instead of having to read stdout (whether by machine or by a user cutting-and-pasting). And I cannot offhand think of a particular reason why that could not just be HEAD (instead of something bisect-specific) after the bisect finishes.

We do not promise any particular value in HEAD now. The only downside would be the minor cost to checkout the working tree of the known-bad commit if we are not already there.

-Peff

On Mon, Jun 24, 2019 at 4:51 AM Jeff King <peff@peff.net> wrote:

It's not clear from he above URL how that differs from what git bisect does. I only found "Returns 0 on success" there which is not very explicit, and we could argue that it's also what git bisect does.

The git bisect documentation says:

       Eventually there will be no more revisions left to inspect, and
the command will print out a description of the first bad commit. The
reference
       refs/bisect/bad will be left pointing at that commit.

So you just need to parse stdout to detect that it found the first bad commit, and then you can use refs/bisect/bad.

If the return code was used, how would you distinguish between a failure in the command (for example if you give bad information to `git bisect good` or `git bisect bad`) and the fact that it has not yet found the first bad commit? Anyway you would need to add some logic for that.

If it would be HEAD, it would mean that git bisect would potentially have to do one more checkout so that HEAD points to the first bad commit. This checkout would sometimes be useless, so it's more efficient to use something like refs/bisect/bad rather than HEAD.

Though we might not explicitly promise in the doc that HEAD will stay at the last commit that was tested, I think that's something people can expect from the way we describe how bisect work. So I don't think it would be a good idea to change our behavior.

On Mon, Jun 24, 2019 at 09:16:13AM +0200, Christian Couder wrote:

Right, thanks for reminding us of that ref. I did leave a suggestion elsewhere in the thread that we might be able to avoid the exit-code pitfalls by using a specific "result" ref.

I'm not too concerned with the cost of checking out a tree. After all, we've just done a bunch of checkouts and probably some operations on the tree itself. But I think there's a much more important reason my suggestion may have problems.

Sometimes we _can't_ checkout another commit, because part of testing the bisection may involve mucking with the working tree. For instance, earlier today I was bisecting a case where git.git did not show the breakage itself, but applying a debugging patch on top did. So I was bisecting along the real history, adding new contents, testing, and then cleaning up my changes before jumping to the next bisection candidate.

On the other hand, I have to clean them up anyway before running "git bisect good" or it cannot jump to the next candidate when we are not finished. Jumping to "bad" when we _are_ finished is no different there. But I think it does show that there can be surprising implications to moving HEAD around.

-Peff

Jeff King <peff@peff.net> writes:

As Christian downthread reminds us, that is what the bisect/bad ref is (which I totally forgot when I gave the earlier response). I do not think we need a new ref, but I do not think it is so bad to add an option "git bisect --exit-code ( --good | --bad ) [<commit-ish>]" that makes the command usually exit with non-zero status. Unless we have found the final answer successfully, that is, and in that case the command would exit with 0 status to signall "all done".

But that should be an option.

On Mon, Jun 24, 2019 at 11:41:53AM -0700, Junio C Hamano wrote:

Ah, right, I forgot that we already have such a ref. So that does seem like a reasonable way to access it (and as Christian notes, it's documented as well). And I agree with you that any exit-code magic should require the user to ask for it explicitly.

I do think that accessing the bisect/bad ref is a little intimate with the implementation (i.e., it implies knowing that there is only a single "bad" that we are moving around, unlike "good", where we may mark many such tips).

So another option is to put the result in a new ref (bisect/result, perhaps), but with one twist: delete it at the beginning of a bisection, so its presence can be used as a marker that the bisection is complete. I.e., you could do "git rev-parse --verify refs/bisect/result" as an alternative to "--exit-code". That gives you the added flexibility of asking "is the bisection done" separately from "do this next bisection step".

-Peff

Jeff King <peff@peff.net> writes:

I actually do not think it is merely an implementation detail, but a rather fundamental part of the "bisect" design. After all, we are trying to find *the* single transition point from "the old world" to "the new world", and the primary assumption that makes "bisection" work in the first place is that there is only one such transition point from old to new. Otherwise, there may be a stretch of history where one transition from "old" to "new" is followed by another transition from "new" back to "old" and then again another one from "old" to "new"; in such a "stripe snake" part of the history, "find the midpoint, if it is from the new world, then narrow to the world before that midpoint", which is the central idea of the "bisection" would not work, would it?

So, ... under that "single transition point" assumption, how would a "bisect" that uses more than one "bad" would work? "bad" is the one that is known to belong to the "new" world that is topologically the oldest and it moves backward in topology to find the boundary from the "old" world. If there were another "bad" (i.e. one that is known to belong the the "new" world, that may not be the oldest), shouldn't that have the other one as its ancestor, making it redundant (i.e. we know we have inherited the badness from the other one)?

Having said that ...

... I kind of like this one exactly for the reasons you give below.