Re: Review of git multimail

[John Keeping <john@keeping.me.uk>]

On Wed, Jul 03, 2013 at 10:02:34AM +0200, Michael Haggerty wrote:
> On 07/03/2013 12:21 AM, Junio C Hamano wrote:
> > Ramkumar Ramachandra <artagnon@gmail.com> writes:
> > 
> >>>     def get(self, name, default=''):
> >>>         try:
> >>>             values = self._split(read_git_output(
> >>>                     ['config', '--get', '--null', '%s.%s' % (self.section, name)],
> >>>                     env=self.env, keepends=True,
> >>>                     ))
> >>
> >> Wait, what is the point of using --null and then splitting by hand
> >> using a poorly-defined static method?  Why not drop the --null and
> >> splitlines() as usual?
> > 
> > You may actually have spotted a bug or misuse of "--get" here.
> > 
> > With this sample configuration:
> > 
> >         $ cat >sample <<\EOF
> >         [a]
> >                 one = value
> >                 one = another
> > 
> >         [b]
> >                 one = "value\nanother"
> >         EOF
> > 
> > A script cannot differentiate between them without using '--null'.
> > 
> > 	$ git config -f sample --get-all a.one
> >         $ git config -f sample --get-all b.one
> > 
> > But that matters only when you use "--get-all", not "--get".  If
> > this method wants to make sure that the user did not misuse a.one
> > as a multi-valued configuration variable, use of "--null --get-all"
> > followed by checking how many items the command gives you back would
> > be a way to do so.
> 
> No, the code in question was a simple sanity check (i.e., mostly a check
> of my own sanity and understanding of "git config" behavior) preceding
> the information-losing next line "return values[0]".  If it had been
> meant as a check that the user hadn't misconfigured the system, then I
> wouldn't have used assert but rather raised a ConfigurationException
> with an explanatory message.
> 
> I would be happy to add the checking that you described, but I didn't
> have the impression that it is the usual convention.  Does code that
> wants a single value from the config usually verify that there is
> one-and-only-one value, or does it typically just do the equivalent of
> "git config --get" and use the returned (effectively the last) value?

Doesn't "git config --get" return an error if there are multiple values?
The answer is apparently "no" - I wrote the text below from
git-config(1) and then checked the behaviour.  This seems to be a
regression in git-config (bisect running now).

I think the "correct" answer is what's below, but it doesn't work like
this in current Git:

    If you want a single value then I think it's normal to just read the
    output of "git config" and let it handle the error cases, without
    needing to split the result at all.

    I think there is a different issue in the "except" block following
    the code quoted at the top though - you will return "default" if a
    key happens to be multi-valued.  The script should check the return
    code and raise a ConfigurationException if it is 2.