[gentoo-dev] Re: [RFC] Perspectives on improving (dis-hacking) python-single-r1

[Luca Barbato <lu_zero@gentoo.org>]

On 23/11/2019 16:48, Michał Górny wrote:
> Hello,
> 
> Some aspects of the current design of python-single-r1 are gross hack.
> I'd like to discuss potential alternatives to them.
> 
> 
> Preamble
> ========
> For the purpose of this mail, let's establish two terms.
> 
> 'Single' will refer to packages allowing the user to select 1 out of N
> supported Python implementations.  This generally means programs linking
> to libpython, foreign build systems and end-user apps where installing
> multiple copies of the code makes no sense.  Those are the packages
> using python-single-r1.
> 
> 'Multi' will refer to packages allowing the user to select M out of N
> supported Python implementations.  This generally means ebuilds for
> Python packages (modules, extensions).  Those are the packages using
> python-r1 (usually via distutils-r1).
> 
> 'Single' packages can PYTHON_USEDEP-end either on other 'single'
> or on 'multi' packages.  However, 'multi' packages can only
> PYTHON_USEDEP-end on other 'multi' packages (because 'single' wouldn't
> ever be able to satisfy >1 impl).
> 
> 
> Current status
> ==============
> Currently, 'single' packages use two sets of USE flags: PYTHON_TARGETS
> and PYTHON_SINGLE_TARGET.  The latter is used to select
> the implementation used, and the former is used only for technical
> reasons.  Selected PST is required to be also enabled in PT.
> 
> All dependencies enforce both PST and PT match.  The PYTHON_USEDEP is
> generally a giant hack that's meant to work both for depending
> on 'single' and 'multi' packages.  I'm not convinced this was the best
> choice anymore.
> 
> In single-to-single deps, PYTHON_USEDEP enforces both PST and PT flags
> to match.  This means that if two packages have matching PST but for
> some reason were built with different PT, users end up having to switch
> PT and rebuild the package without any real change.
> 
> In single-to-multi deps, PYTHON_USEDEP enforces PT flags.  For this
> reason, we need to enforce that selected PST is always present in PT,
> and users always have to put both flags in package.use.
> 
> There's also a gross hack in PYTHON_USEDEP in 'multi' packages that
> means to prevent multi-to-single deps.  However, the dep mismatch it
> causes is not very readable.
> 
> 
> Alternative 1: PYTHON_TARGETS only
> ==================================
> The first alternative I'd like to explore is removing
> PYTHON_SINGLE_TARGET flags.  Why were they added in the first place?
> The primary goal was to be able to set PT to 2.7+3.x without requiring
> people to adjust flags for every 'single' package out there.  Plus, it
> turned out very convenient for Python team members who want to enable
> all PT but obviously can't do the same for PST.
> 
> Originally I brought this proposal in context of automated REQUIRED_USE
> conflict resolution.  However, that was rejected by the Council.  Still,
> it may start making sense again in the near future when we start
> removing py2.7 and pypy2.7 support.  If we can limit PT to one
> implementation, and handle the remaining packages via IUSE defaults
> and package.use, this may just work.
> 
> The inconvenience to people enabling multiple PT will remain though.
> 
> 
> Alternative 2: new dependency API
> =================================
> If PST is going to stay, we may look into removing PT from 'single'
> packages instead.  The idea is to provide new method of generating
> cross-package deps that doesn't require fake flags.
> 
> PYTHON_USEDEP would continue working through a transitional period.
> When it's entirely gone, we can remove PYTHON_TARGETS from 'single'
> packages.
> 
> Single-to-single deps would switch to PYTHON_SINGLE_USEDEP, that only
> enforces PST and disregards PT entirely.
> 
> Single-to-multi deps would have to use a new generator function,
> $(python_gen_multi_dep ...) that would create appropriate USE-mapping
> from PST to PT.
> 
> Example ebuild would have:
> 
> RDEPEND="app-foo/singlepkg[${PYTHON_SINGLE_USEDEP}]
>      $(python_gen_multi_dep '
>           dev-python/foo[${PYTHON_USEDEP}]
>           dev-python/bar[${PYTHON_USEDEP}]
>      ')"
> 
> This will generate something like:
> 
>      pst_python2_7? (
>          dev-python/foo[pt_python2_7(-)]
>          dev-python/bar[pt_python2_7(-)]
>      )
>      pst_python3_7? (
>          dev-python/foo[pt_python3_7(-)]
>          dev-python/bar[pt_python3_7(-)]
>      )
> 
> 
> Your opinions
> =============
> So, WDYT?  Do you think this approach is worthwhile?  Do you see other
> options?
> 
> 

The new dependency API proposal looks nicer to me.

lu