[R-pkg-devel] Handling Not-Always-Needed Dependencies?
Thomas J. Leeper
thosjleeper at gmail.com
Wed Aug 3 15:20:59 CEST 2016
The issue seems to boil down to the fact that Suggests is covering two
very different use cases: (a) conditional package code and (b)
example/test/vignette code.
Consider a package (say "foo") that is only used in tests. Users do
not need "foo" since package code never calls it. If our package
instead calls "foo" conditionally (requireNamespace(); foo::bar(),
etc.), then users may very well want "foo" and need it much more than
they would if "foo" were only used in tests. Yet DESCRIPTION does not
allow a distinction two be made between these two scenarios.
I think the length and complexity of the discussion in WRE[1] makes
clear that Suggests is being used two cover these two very different
use cases.
-Thomas
[1] https://cran.r-project.org/doc/manuals/r-devel/R-exts.html#Package-Dependencies
Thomas J. Leeper
http://www.thomasleeper.com
On Wed, Aug 3, 2016 at 2:04 PM, Mark van der Loo
<mark.vanderloo at gmail.com> wrote:
>>> Recommends: only gets installed, can be used via if(requireNamespace())
>>> from the package and in pkg tests[1] [snip]
>>> Build-Depends: gets installed before build, removed after.
>>> Suggests: only installed when requested at install.packages() and only
>>> used in examples/vignettes.
>
> [snip]
>> I'd much rather
>> have a way of declaring explicitly the different aspects of dependence
>> on a package rather than bundling them up into cute labels,
>
> Agreed
>
>> but it's too late for that now. However, we don't need to make things
>> worse.
>
> Disagreed. We could follow the well-established practices of Debian (and
> CRAN already does that, partially).
>
>
>>> If 'tons of packages' are using if(requireNamespace) in their package
>>> code there seems to be a need for something like this. [snip]
>
>> I don't follow the argument here. What problem are you solving?
>
> Basically I'm trying to address the idea suggested by Thomas, who started
> this conversation, and make it a bit more explicit. I felt that the
> discussion went a little off-track there.
>
> Right now, when package code (not examples) uses a suggested package, part
> of that package will by default not work - at least that's how people use it
> now. I would like it to work by default. For examples/vignettes you could be
> more forgiving since running an example is not core functionality of a
> package.
>
>>> Perhaps more controversially a 'Breaks' field could be considered. [snip]
>
>> This isn't controversial, it's just a bad idea. Don't encourage people
>> to break things.
>
> Your reaction just proved my point about it being controversial. More
> seriously, real progress is hardly ever possible without breaking things, so
> I think at least people could have a serious discussion about it before
> dismissing it simply as a bad idea. The Debian community obviously once
> thought it was a good idea, so why not discuss it for R/CRAN? (discussions
> are also an important way to progress even if no line of code is changed).
> At the moment, I'm inclined against the idea, but I for one like to see me
> proven wrong.
>
>
>>> [1] actually, once we know a pkg is Recommended, the
>>> 'if(requireNamespace)' could even be absorbed in the :: operator.
>
>>I don't see how :: would be any different than it is now. If you don't
>>have foo available, and you try to use foo::bar(), what would happen
>>other than an error?
>
> I think you're right there. <resets brain>.
>
> Best,
> Mark
>
>
> Op wo 3 aug. 2016 om 13:41 schreef Duncan Murdoch
> <murdoch.duncan at gmail.com>:
>>
>> On 03/08/2016 5:32 AM, Mark van der Loo wrote:
>> >
>> > After reading the link in Dirk's initial reply, how about adding fields
>> > 'Recommends' and 'Build-Depends' to DESCRIPTION as in Debian?
>> >
>> > Recommends: only gets installed, can be used via if(requireNamespace())
>> > from the package and in pkg tests[1]. [Debian: The Recommends field
>> > should list packages that would be found together with this one in all
>> > but unusual installations.]
>> > Build-Depends: gets installed before build, removed after.
>> > Suggests: only installed when requested at install.packages() and only
>> > used in examples/vignettes.
>>
>> I think the distinction between Recommends and Suggests is too subtle
>> here. I already think it's a bad thing that we are using these words in
>> ways that don't really correspond to English usage. I'd much rather
>> have a way of declaring explicitly the different aspects of dependence
>> on a package rather than bundling them up into cute labels, but it's too
>> late for that now. However, we don't need to make things worse.
>>
>> >
>> > If 'tons of packages' are using if(requireNamespace) in their package
>> > code there seems to be a need for something like this. Compliance to the
>> > above can be checked automatically and a gradual implementation via
>> > NOTE->WARNING->ERROR in R CMD check seems possible.
>>
>> I don't follow the argument here. What problem are you solving?
>>
>> > Perhaps more controversially a 'Breaks' field could be considered. There
>> > are a few packages out there that have many, many, dependencies.
>> > Implementing breaking updates currently depends on the willingness of
>> > many authors to update their package or convincing the CRAN maintainers
>> > to allow for (temporary) breakage.
>>
>> This isn't controversial, it's just a bad idea. Don't encourage people
>> to break things.
>>
>> > The suggestion to have functions auto-install things is very
>> > inconvenient for the good reasons pointed out by Thomas. Additionally,
>> > it is often based on the wrong assumptions. Example: the RGtk2 package
>> > has this habit of trying to install when libgtk2 is not on the path. But
>> > in my case that is often exactly the case: it is just not on the path
>> > (libgtk2 is on the network, the VM just doesn't know yet). So I'd rather
>> > have a proper and accurate error message (which is good practice
>> > anyway).
>> >
>> >
>> > Best,
>> > Mark
>> >
>> > [1] actually, once we know a pkg is Recommended, the
>> > 'if(requireNamespace)' could even be absorbed in the :: operator.
>>
>> I don't see how :: would be any different than it is now. If you don't
>> have foo available, and you try to use foo::bar(), what would happen
>> other than an error?
>>
>> Duncan Murdoch
>>
>> >
>> >
>> >
>> >
>> > Op wo 3 aug. 2016 om 01:46 schreef Duncan Murdoch
>> > <murdoch.duncan at gmail.com <mailto:murdoch.duncan at gmail.com>>:
>> >
>> > On 02/08/2016 6:34 PM, Dirk Eddelbuettel wrote:
>> > >
>> > > On 2 August 2016 at 18:13, Duncan Murdoch wrote:
>> > > | Okay, now I think I understand, but I agree with CRAN. It is
>> > not
>> > > | feasible to tell if the test happened somewhere in the code
>> > unless we
>> > > | enforce a particular way of writing the test.
>> > >
>> > > Debian has well over 20k packages, and they are tested this way.
>> > You just
>> > > need to show the will of testing in an _empty_ environment to
>> > ensure
>> > > _everything_ that is needed is loaded.
>> > >
>> > > | I would object if I had to write if (requireNamespace("foo"))
>> > multiple
>> > > | times just to satisfy CRAN's test, when any sane human could
>> > tell that
>> > > | the first test was sufficient.
>> > > |
>> > > | For example, if my package Suggests: foo, I should be able to
>> > write
>> > > |
>> > > | if (!requireNamespace("foo"))
>> > > | stop("Package foo is needed for this example")
>> > > |
>> > > | and then merrily call foo::bar() as many times as I like.
>> > > |
>> > > | Or am I still misunderstanding you? What particular thing
>> > should CRAN
>> > > | change?
>> > >
>> > > You seem to misunderstand that both you and I want
>> > >
>> > > if (!requireNamespace("foo"))
>> > > stop("Package foo is needed for this example")
>> > >
>> > > (or alternative per-call tests) and that CRAN does not enforce
>> > either.
>> > >
>> > > CRAN, like Hadley, just closes its eyes, swallows hard, and then
>> > simply loads
>> > > everything treating Suggests as if it were Depends.
>> > >
>> > > But it ain't. Suggests != Depends.
>> > >
>> > > Now clearer?
>> >
>> >
>> > So really what you're suggesting is that CRAN should run tests with
>> > the
>> > suggested packages absent. Presumably tests should also be run with
>> > them present.
>> >
>> > But if they did that, the code that I want to write would call
>> > stop()
>> > and fail. So we'd need some way to say "Let the user know they need
>> > 'foo' to run this, but don't fail." And we'd need to phase this in
>> > really gradually, because tons of packages are using code like my
>> > example.
>> >
>> > You volunteered to help CRAN package checking. Why not put together
>> > code to implement your idea, and see how big the problem would be to
>> > phase it in, by seeing how many packages fail under it?
>> >
>> > Duncan Murdoch
>> >
>> > ______________________________________________
>> > R-package-devel at r-project.org <mailto:R-package-devel at r-project.org>
>> > mailing list
>> > https://stat.ethz.ch/mailman/listinfo/r-package-devel
>> >
>>
>
More information about the R-package-devel
mailing list