[Distutils] Fwd: Re: PEP 517 again

[Paul Moore]

On 24 August 2017 at 15:26, Thomas Kluyver <thomas at kluyver.me.uk> wrote:
> I particularly want to hear from Nick, Daniel, Donald and Paul on their
> understanding of the build_directory parameter. There were reasons why it
> got added, and even if those reasons no longer seem so convincing to me,
> it's only going to be removed again if we can reach a consensus on working
> without it.

OK. I've not had much time recently (RL issues) and so my apologies if
I've forgotten context here. But my feelings are:

1. I've been pretty much persuaded that including features in the
interface to defend against badly written backends is not worthwhile.
(I still worry that it may be hard to provide good error reporting for
backend errors, but that's a lesser issue).
2. I'm not completely clear on how pip's implementation will work - I
think the intention is to always build a sdist and build a wheel from
that, unless the backend reports it can't build a sdist, in which case
we ask it to build a wheel directly.
3. I'm not particularly comfortable with backends saying that they can
only produce a wheel and not a sdist, but accept that might have to be
a reality in some cases. I would be *really* bothered by a backend
that wasn't even willing to accept that sdist->wheel and direct wheel
builds should be equivalent - but I don't know if that's something
that is likely to happen (we can of course state that expectation in
the PEP). The PEP seems to think this is possible, though[1] (see my
next point).
4. The point of all this is that the definition of build_directory
says "If build_directory is None, the backend may do an 'in place'
build which modifies the source directory and may produce different
results from those that would be obtained by exporting an sdist
first". That's the bit that bothers me a lot - if we lose
build_directory, we lose the option for pip to request that a backend
produce a wheel "equivalent to producing an sdist first". And that
capability is something I think pip should require[2]. I don't
particularly care who specifies the actual directory (as per your
comment, I thought someone else wanted this, for caching or something)
but I *do* care about telling backends we don't want "different
results from those that would be obtained by exporting an sdist
first".

Paul

[1] The only example I can think of is flit where the user doesn't
have git installed. But then I'd expect that a wheel generated
directly in that case would be the same as if the user installed git
then did sdist->wheel, so I'd still consider that to be maintaining
the invariant I care about.
[2] I'm not likely to be the one implementing pip's PEP 517 support,
so I could be overridden on this, of course.