Re: .gitattributes override behavior (possible bug, or documentation bug)

From: Dakota Hawkins <dakota@dakotahawkins.com>
Date: Mar 20, 2018, 03:10 UTC
Message-ID: <CAHnyXxRcwq40W4tKm=Kscrsnb77yh7=eGDE=r5AZq073MPX9AQ@mail.gmail.com>
In-Reply-To: <20180320023423.GA10143@sigill.intra.peff.net>

Thanks for the quick reply!

On Mon, Mar 19, 2018 at 10:34 PM, Jeff King <peff@peff.net> wrote:

Show 19 quoted lines

> On Mon, Mar 19, 2018 at 09:49:28PM -0400, Dakota Hawkins wrote:
>
>> Summary: Trying to apply attributes to file extensions everywhere
>> except in one directory.
>>
>> .gitattributes:
>>
>>     *.[Pp][Nn][Gg] filter=lfs diff=lfs merge=lfs -text
>>     /.readme-docs/ -filter=lfs -diff=lfs -merge=lfs
>>
>> Make some data:
>>
>>     echo "asldkjfa;sldkjf;alsdjf" > ./.readme-docs/test.png
>>     git add -A
>
> As you noted below, that second line does not match your path, because
> attributes on a directory aren't applied recursively. And it has nothing
> to do with overriding. If you remove the png line entirely, you can see
> that we still do not match it. You need to use "*" to match the paths.

Ah, yes, I see that. Inconsistent with .gitignore (more below), right?

Show 22 quoted lines

> You may also find that "-diff=lfs" does not do quite what you want.
> There is no way to say "cancel any previous attribute", which I think is
> what you're trying for here. You can only override it with a new value.
> So:
>
>   /.readme-docs/* -diff
>
> says "do not diff this". And:
>
>   /.readme-docs/* diff
>
> says "diff this as text, even if it looks binary".
>
> The best you can probably do is:
>
>   /.readme-docs/* diff=foo
>
> Since you have no diff.foo.* config, that will behave in the default way
> (including respecting the usual "is it binary" checks). So a bit hacky,
> but I think it would work as "ignore prior diff".
>
> And I think filter and merge drivers should work the same.

That's interesting... in this case I was taking my advice on how this should work from the git-lfs folks. I have promised to share what I find here with them, so that will help at least :)

I think that makes sense to me -- there would be no good way to tell it what the default should have been without explicitly telling it what to use instead.

Show 16 quoted lines

>> Is this me misunderstanding something in the documentation? I would
>> expect "./.readme-docs/" to match "./.readme-docs/test.png" and
>> override the earlier "*.[Pp][Nn][Gg]" attributes.
>>
>> I have found the following overrides to work in lieu of the directory match:
>>
>>     /.readme-docs/* -filter=lfs -diff=lfs -merge=lfs
>>     /.readme-docs/**/* -filter=lfs -diff=lfs -merge=lfs
>>
>> ...but I don't see a justification in the documentation for this
>> working and the original directory filter not working.
>
> I could not find anything useful in gitattributes(5). There's some old
> discussion here:
>
>   https://public-inbox.org/git/slrnkldd3g.1l4.jan@majutsushi.net/

If I follow that correctly: There's some initial speculation that it would be OK to apply the attributes recursively, which is then shot down because it wasn't designed to be recursive (though I don't see a different, technical reason for that), followed by finding a (this same?) solution/workaround for the original problem. Is that about right?

> which makes it clear that attributes aren't recursive, but it's probably
> worth calling out in the documentation. In fact, I think the current
> documentation is a bit misleading in that it says "patterns are matched
> as in .gitignore", which is clearly not the case here.

I was indeed going off of the suggestion to consult the .gitignore pattern matching documentation.

> I think just "/.readme-docs/**" should be sufficient for your case. You
> could also probably write "*" inside ".readme-docs/.gitattributes",
> which may be simpler (you don't need "**" there because patterns without
> a slash are just matched directly against the basename).

Wouldn't that make the "*" inside ".readme-docs/.gitattributes", technically recursive when "*" matches a directory? It's always seemed to me that both were necessary to explicitly match things in a directory and its subdirectories (example, IIRC: "git ls-files -- .gitattributes" vs "git ls-files -- .gitattributes **/.gitattributes"). Maybe that example is peculiar in that its a dotfile and can't have a wildcard before the dot?

I guess my takeaway is that it would be _good_ if the gitattributes documentation contained the caveat about not matching directories recursively, but _great_ if gitattributes and gitignore (and whatever else there is) were consistent.

At any rate, thanks for the great, quick help!

-Dakota