Re: Bizarre missing changes (git bug?)

From: Tim Harper <timcharper@gmail.com>
Date: Jul 21, 2008, 22:53 UTC
Message-ID: <E301C92A-8794-4E90-9C85-D73B94A2648C@gmail.com>
In-Reply-To: <alpine.LFD.1.10.0807211331390.31863@woody.linux-foundation.org>

On Jul 21, 2008, at 2:37 PM, Linus Torvalds wrote:

Show 33 quoted lines

>
>
> On Mon, 21 Jul 2008, Tim Harper wrote:
>>
>> Anyone run into this before?  Any idea what might have caused it?   
>> We're a bit
>> concerned about this because if we don't know how to avoid this, we  
>> no longer
>> can feel certain that when something is committed, it will make it  
>> out in our
>> release.
>
> Read up on '--full-history'.
>
> By default, git simplifies the history for logs that have path
> simplification: only walking down the side of a merge that all the  
> data
> came from (ie the unchanged side). So it only leaves merges around if
> there was changes from _all_ parents.
>
> So without --full-history, if any parent matches the state, it just
> removes the merge and picks that parent that contained all the state.
> Obviously, any changes to that file can be sufficiently explained by
> walking just that limited history, because they must have changed in
> _that_ history too!
>
> That default behaviour leads to a *much* simpler history, and is  
> usually
> what you want - it avoids unnecessary duplication when something was
> changed trivially the same way in both branches - 'git log' will  
> just pick
> the first branch.
>

Agreed - this was an insightful decision.

Show 12 quoted lines

> So, if you had two (or more) commits that both fixed the same bug in
> different branches, and thus both branches actually ended up with  
> the same
> contents, it does mean that "git log <filename>" will only show  
> _one_ of
> the fixes.
>
> In this case, it apparently showed another version than the one you  
> were
> looking for.
>
> 				Linus

Git has made me feel stupid on various occasions. This is no exception as the problem turned out being in the chair, not in git.

After running through git bisect, and ran the command Alex Riesen suggested, it made it pretty crystal clear where things went wrong. It turned out to be a bad merge that was from a conflict related to white-space issues, and the wrong resolution was chosen (a resolution that also consequently turned out to be no change).

Another false impression I had is a merge conflict resolution would always be displayed in a merge commit. However, after running over the merges again, if you pick the right or left, discarding the one or the other, nothing is shown in "git log -p" for the merge commit. Is there a way to see what was chosen for a conflict resolution? Seeing that in the merge commit would have made things a little more clear.

Thank you for articulating git branch's behavior - all is clear as mud now :)

Tim