[Distutils] Provisionally accepting PEP 517's declarative build system interface

Donald Stufft donald at stufft.io
Sat Jun 3 03:47:26 EDT 2017

> On Jun 3, 2017, at 1:40 AM, Nathaniel Smith <njs at pobox.com> wrote:
> On Fri, Jun 2, 2017 at 8:38 PM, Donald Stufft <donald at stufft.io> wrote:
>> On Jun 2, 2017, at 10:14 PM, Nathaniel Smith <njs at pobox.com> wrote:
>>> So far my belief is that packages with expensive build processes are
>>> going to ignore you and implement, ship, document, and recommend the
>>> direct source-tree->wheel path for developer builds. You can force the
>>> make-a-wheel-from-a-directory-without-copying-and-then-install-it
>>> command have a name that doesn't start with "pip", but it's still
>>> going to exist and be used. Why wouldn't it? It's trivial to implement
>>> and it works, and I haven't heard any alternative proposals that have
>>> either of those properties. [1]
>> If someone wants to implement a direct-to-wheel build tool and have it
>> compete with ``pip install .`` they’re more than welcome to. Competition is
>> healthy and at the very worst case it could validate either the idea that
>> direct-to-wheel is important enough that people will gladly overcome the
>> relatively small barrier of having to install another tool and then we have
>> data to indicate maybe we need to rethink things or it could validate the
>> idea that it’s not important enough and leave things as they are.
>> I went and looked through all 105 pages of pip’s issues (open and closed)
>> and made several searches using any keyword I could think of looking for any
>> issue where someone asked for this. The only times I can find anyone asking
>> for this were you and Ralf Gommers as part of the extended discussion around
>> this set of PEPs and I’ve not been able to find a single other person asking
>> for it or complaining about it.
> That's because until now, the message that everyone has received over
> and over is that the way you install a package from a directory on
> disk is:
>  cd directory
>  python setup.py install
> and this does incremental builds. (My experience is that even today,
> most people are surprised to learn that 'pip install' accepts
> directory paths.)

I’m not sure this is true? I mean I’m sure it’s true that for *some* people that’s the message they get, but we see a reasonable volume of bug reports and the like from people passing a path to pip that it’s not hardly an un(der)used feature in the slightest. We have zero metrics so I suspect there is no way to answer which one is used more than the other though. I think writing either one off as “nobody uses that method” is likely to be the wrong answer.

Pip is generally used widely enough in enough different scenarios that it is unusual for any feature it has (even ones that are completely undocumented and requires diving in the source code!) to not be used by a decent chunk of people. I’m not saying that zero people exist that would want this (indeed, there is at least two!) but that it has obviously not been pressing enough to need it that someone felt the need to open a ticket or ask for it before.

> In our glorious PEP 517 future, we have to teach everyone to stop
> using 'setup.py install' and instead use 'pip install .'. This switch
> enables a glorious new future of non-distutils-based build systems and
> fixes a bunch of other brokenness at the same time, hooray, BUT
> currently switching to 'pip install' also causes a regression for
> everyone who's used to incremental builds working.
> Ralf and I noticed this because we were looking at getting a head
> start on the glorious future by making 'pip install' mandatory for
> numpy and scipy. The reason no-one else has noticed is that we're
> among the few people that have tried using 'pip install' as their
> standard install-from-working-tree command. But soon there will be
> more.

One thing I’d note is that as far as I can tell, neither the current copy of the PEP, nor my PR or Thomas’ PR or any of the discussions here have talked about having the interface itself mandate calling build_sdist prior to calling build_wheel. I’m personally fine if we want the interface to allow it (and I assumed we would TBH) and explicitly calling that out.

That means that whether or not you call build_sdist (or some copy files hook) first effectively ends up being an implementation detail of the frontend in question. That would allow pip to build the sdist (or do the copy thing) first but another tool could implement it the other way (and installing a local wheel is so simple that you could pretty easily implement that and just shell out to pip for the rest if you wanted).

That also means that we can adjust our answer to it in the future. If such a tool gets built and a lot of people end up using it and asking for it in pip, we can revisit that decision in a future version of pip. Part of the stand off here is the pip developers view it as a regression if we stop building in isolation and you view it as a regression if incremental/inplace builds are not supported. Both can be true! It’s the opinion of the pip developers who have spoken so far that for us, the risk of our regressions is high enough we don’t currently feel comfortable changing that behavior. 

>> However, what I was able to find was what appears to be the original reason
>> pip started copying the directory to begin with,
>> https://github.com/pypa/pip/issues/178 which was caused by the build system
>> reusing the build directory between two different virtual environments and
>> causing an invalid installation to happen. The ticket is old enough that I
>> can get at specifics it because it was migrated over from bitbucket. However
>> the fact that we *used* to do exactly what you want and it caused exactly
>> one of problem I was worried about seems to suggest to me that pip is
>> absolutely correct in keeping this behavior.
> Hmm, it looks to me like that bug is saying that at the time, if you
> ran 'python setup.py install' *inside the pip source tree*, and then
> tried to run pip's test suite (possibly via 'setup.py test'), then it
> broke. I don't think this is related to the behavior of 'pip install
> .', and I feel like we would know if it were currently true that
> running 'setup.py install' twice in the same directory produced broken
> shebang lines. (Again, most people who install from source directories
> are currently using setup.py install!)

Meh yea, I misread the bug report. Reading 105 pages of reports will do that I guess :P

> The source tree copying was originally added in:
>   https://github.com/pypa/pip/commit/57bd8163e4483b7138342da93f5f6bb8460f0e4a
> (which is dated ~2 months before that bug you found, and if I'm
> reading it right tweaks a code path that previously only worked for
> 'pip install foo.zip' so it also works for 'pip install foo/'). AFAICT
> the reason it was written this way is that pip started out with the
> assumption that it was always going to be downloading and unpacking
> archives, so the logic went:
> 1) make a temporary directory
> 2) unpack the sdist into this temporary directory
> 3) build from this temporary directory
> Then, when it came time to add support for building from directories,
> the structure of the logic meant that by the time pip got to step (2)
> and realized that it already had a source directory, it was too late
> -- it was already committed to using the selected temporary directory.
> So instead of refactoring all this code, they made the minimal change
> of implementing the "unpack this sdist into this directory" operation
> for source directories by using shutil.copytree.
> I think this chain of reasoning will feel very familiar to anyone
> working with the modern pip source 5 years later...
> It's absolutely true that there are cases where incremental builds can
> screw things up, especially when using distutils/setuptools. But I
> don't think this is why pip does things this way originally :-).
>> It’s not that I don’t trust the backend, it’s that I believe in putting in
>> systems that make it harder to do the wrong thing than the right thing. As
>> it is now building in place correctly requires the build backend to do extra
>> work to ensure that some file that wouldn’t be included in the sdist doesn’t
>> influence the build in some way. Given that I’m pretty sure literally every
>> build tool in existence for Python currently fails this test, I think that
>> is a pretty reasonable statement to say that it might continue to be a
>> problem into the future.
>> Copying the files makes that harder to do (but still easier than always
>> going through the sdist). If you want to argue that we should always go
>> through the sdist and we shouldn’t have a copy_files hook, I’m ok with that.
>> I’m only partially in favor of it as a performance trade off because I think
>> it passes a high enough bar that it’s unlikely enough for mistakes to be
>> made (and when they do, they’ll be more obvious).
> What do you think of letting build backends opt-in to in-place builds?

I think the PEP already allows the interface to be used for in-place builds, though it could use some language specifying that and making sure it’s explicit that is an option the PEP allows. I think I’m neutral on allowing backends to express a preference for in-place builds, but again I don’t think I would be comfortable listening to that to allow a backend to do an in-place build, at least not by default or at first. Could I see us get to a place where we did allow that? Maybe. I wouldn’t promise anything one way or another (and TBH how pip chooses to implement a PEP isn’t really something that the PEP or distutils-sig gets to choose) but I’m fine with leaving the mechanisms in place to allow that in the future.

>>> Other unresolved issues:
>>> - Donald had some concerns about get_wheel_metadata and they've led to
>>> several suggestions, none of which has made everyone go "oh yeah
>>> obviously that's the solution". To me this suggests we should go ahead
>>> and drop it from PEP 517 and add it back later if/when the need is
>>> more obvious. It's optional anyway, so adding it later doesn't hurt
>>> anything.
>> My main concern is the metadata diverging between the get_wheel_metadata and
>> the building wheel phase. The current PEP solves that in a reasonable enough
>> way (and in a way I can assert against). My other concerns are mostly just
>> little API niggles to make it harder to mess up.
>> I think this one is important to support because we do not to be able to get
>> at the dependencies, and invoking the entire build chain to do that seems
>> like it will be extraordinarily slow.
> It's only slow in the case where (a) there's no wheel (obviously), and
> (b) after getting the dependencies we decide we don't want to install
> this sdist after all. I imagine numpy for example won't bother
> implementing get_wheel_metadata because we provide wheels for all the
> platforms we support and because we have no dependencies, so it is
> doubly useless AFAICT. But yeah in other cases it could matter. I'm
> not opposed to including it in general, just thought this might be a
> way to help get the minimal PEP 517 out the door.

Yea I don’t think this is really a sticking point, I think it’s mostly just around:

1) Do we include a build_sdist hook?
    A) Can pip use this as part of the process of going from a VCS to a wheel
2) Do we include a copy_the_files hook?
3) Minor decisions like unpacked vs packed wheels/sdists and cwd vs pass in a path.

>>> - It sounds like there's some real question about how exactly a build
>>> frontend should handle the output from build_wheel; in particular, the
>>> PEP should say what happens if there are multiple files deposited into
>>> the output dir. My original idea when writing the PEP was that the
>>> build frontend would know the name/version of the wheel it was looking
>>> for, and so it would ignore any other files found in the output dir,
>>> which would be forward compatible with a future PEP allowing
>>> build_wheel to drop multiple wheels into the output dir (i.e., old
>>> pip's would just ignore them). It's clear from the discussion that
>>> this isn't how others were imagining it. Which is fine, I don't think
>>> this is a huge problem, but we should nail it down so we're not
>>> surprised later.
>> How do you determine the name/version for ``pip install .`` except by
>> running get_wheel_metadata or build_wheel or build_sdist?
> Well, I was imagining that the semantics of 'pip install .' in a
> multi-wheel world would be to install all the generated wheels :-).
> But yeah, it's not really well-specified as currently written.
> Possibly the simplest solution is to say that build_wheel has to
> return a string which names the wheel, and then in the future we could
> add build_wheel2 which is identical but returns a list of strings, and
> backwards compatibility would be:
> def build_wheel2(...):
>    return build_wheel(...)[0]

That seems reasonable to me.

>> -n
>>> [1] Donald's suggestion of silently caching intermediate files in some
>>> global cache dir is unreasonably difficult to implement in a
>>> user-friendly way – cache management is Hard, and I frankly I still
>>> don't think users will accept individual package's build systems
>>> leaving hundreds of megabytes of random gunk inside hidden
>>> directories. We could debate the details here, but basically, if this
>>> were a great idea to do by default, then surely one of
>>> cmake/autoconf/... would already do it? Also, my understanding is the
>>> main reason pip wants to copy files in the first place is to avoid
>>> accidental pollution between different builds using the same local
>>> tree; but if a build system implements a global cache like this then
>>> surprise, now you can get pollution between arbitrary builds using
>>> different trees, or between builds that don't even use a local tree at
>>> all (e.g. running 'pip install numpy==1.12.0' can potentially cause a
>>> later run of 'pip install numpy==1.12.1' to be corrupted). And, it
>>> assumes that all build systems can easily support out-of-tree
>>> incremental builds, which is often true but not guaranteed when your
>>> wheel build has to wrap some random third party C library's build
>>> system.
>> Make it opt-in
> If it's opt-in, then I might as well tell people to run 'pip
> devinstall .' or 'in-place-install .' or whatever instead, and it'll
> be much easier all around. But instead of making it opt-in, I'd much
> rather it Just Work. It's frustrating that at the same time we're
> moving to the glorious simplified future, we're also picking up a new
> piece of arcane wisdom that devs will need to be taught, and another
> place where numerical Python devs will roll their eyes at how the
> standard Python tooling doesn't care about them. (And I totally
> understand that the motivation on your end is also to make things Just
> Work, but I feel like in the specific case where someone is
> *repeatedly* building out of the *same source directory* – which is
> the one in dispute here – we should optimize for developer
> experience.)

People repeatably build out of the same source directory for lots of different reasons. I just closed a ticket the other day about someone who was trying to do a ``-e .`` in a read only directory because they had it mounted inside of a container or a VM or so and were editing it from another machine and it didn’t work (because -e obviously builds in place).

One more specific example I’m concerned about with regressions here is a case like:

docker run -it ubuntu:latest -v $PWD:/app/ pip install /app/
docker run -it centos:latest -v $PWD:/app/ pip install /app/

I’ve seen people doing similar things in the wild today, and if an in place build directory just uses the normal platform tuples to differentiate build directories, there is a good chance that is going to fail miserably, likely with some sort of god awful linking error or segfault.

Donald Stufft

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/distutils-sig/attachments/20170603/983e2ede/attachment-0001.html>

More information about the Distutils-SIG mailing list