[Distutils] Reproducible builds (Sdist)

[Nick Coghlan]

On 30 September 2017 at 06:02, Thomas Kluyver <thomas at kluyver.me.uk> wrote:
> On Fri, Sep 29, 2017, at 07:16 PM, Matthias Bussonnier wrote:
>> Second; is there a convention to store the SDE value ? I don't seem to
>> be able to find one. It is nice to have reproducible build; but if
>> it's a pain for reproducers to find the SDE value that highly decrease
>> the value of SDE build.
>
> Does it make sense to add a new optional metadata field to store the
> value of SOURCE_DATE_EPOCH if it's set when a distribution is built? I
> guess it could cause problems if unpacking & repacking a tarball means
> that its metadata is no longer accurate, though.

For distro level reproducible build purposes, we typically treat the
published tarball *as* the original sources, and don't really worry
about the question of "Can we reproduce that tarball, from that VCS
tree?".

This stems from the original model of open source distribution, where
publication *was* a matter of putting a tarball up on a website
somewhere, and it was an open question as to whether or not the
publisher was even using a version control system at all (timeline:
RCS=1982, CVS=1986, SVN=2000, git/hg=2005, with Linux distributions
getting their start in the early-to-mid 1990's).

So SOURCE_DATE_EPOCH gets applied *after* unpacking the original
tarball, rather than being used to *create* the tarball (we already
know when the publisher created it, since that's part of the tarball
metadata).

Python's sdists mess with that assumption a bit, since it's fairly
common to include generated C files that aren't part of the original
source tree, and Cython explicitly recommends doing so in order to
avoid requiring Cython as a build time dependency:
http://docs.cython.org/en/latest/src/reference/compilation.html#distributing-cython-modules

So in many ways, this isn't the problem that SOURCE_DATE_EPOCH on its
own is designed to solve - instead, it's asking the question of "How
do I handle the case where my nominal source archive is itself a built
artifact?", which means you not only need to record source timestamps
of the original inputs you used to build the artifact (which the
version control system will give you), you also need to record details
of the build tools used (e.g. using a different version of Cython will
generate different code, and hence different "source" archives), and
decide what to do with any timestamps on the *output* artifacts you
generate (e.g. you may decide to force them to match the commit date
from the VCS).

So saying "SOURCE_DATE_EPOCH will be set to the VCS commit date when
creating an sdist" would be a reasonable thing for an sdist creation
tool to decide to do, and combined with something like `Pipfile.lock`
in `pipenv`, or a `dev-requirements.txt` with fully pinned versions,
*would* go a long way towards giving you reproducible sdist archives.

However, it's not a problem to be solved by adding anything to the
produced sdist: it's a property of the publishing tools that create
sdists to aim to ensure that given the same inputs, on a different
machine, at a different time, you will nevertheless still get the same
result.

Cheers,
Nick.

-- 
Nick Coghlan   |   ncoghlan at gmail.com   |   Brisbane, Australia