[Rd] How graphical an interface should the default be?

[Prof Brian Ripley]

On Fri, 25 Feb 2005, Robert Gentleman wrote:

> On Feb 25, 2005, at 3:11 AM, Prof Brian Ripley wrote:
>
>> install/update.packages will have a lot of changes in 2.1.0, and I have 
>> been adding some widgets to go along with this.
>> 
>> - Rather than just CRAN and BIOC, you have a character vector of
>>   repositories.  There is a function setRepositories() to set the
>>   appropriate option().
>> 
>> - There is no default CRAN, but a function chooseCRANmirror() to set a
>>   mirror, which is invoked if you try to access CRAN without setting a
>>   mirror.
>> 
>> - update.packages(ask="graphics") brings up a listbox for you to de-select
>>   packages (all available updates are pre-selected).
>> 
>> - install.packages() with no/empty pkgs argument brings up a listbox of
>>   all available packages (including those inside bundles).
>> 
>> - menu(graphics=TRUE) is implemented.
>> 
>> These can be set up to use widgets where available (Windows, if Tk is 
>> available under X11, I hope Aqua before release), and have a text-mode 
>> fallback (better than the current menu(), but in that spirit). (Except that 
>> is install.packages: text-mode selection from 480 packages even in three 
>> columns is not useful to me, but a scrolling list works well as Windows 
>> users of R already know.)
>> 
>> My question is:
>> 
>> 	What should be default be?
>> 
>> Options might be:
>> 
>> - use the graphics widget if available.
>> 
>> - make the graphics the default on Windows (it will always be available,
>>   but not necessarily on the right screen under DCOM uses).
>> 
>> - make text-mode the default on Unix.
>> 
>> - do different things for different tools.  Currently setRepositories()
>>   and chooseCRANmirror() default to graphics-if-possible, and
>>   update.packages() defaults to text mode (which gives more
>>   detailed information).
>> 
>> I am not really in favour of making the defaults a set of options, but that 
>> is possible.
>> 
>> 
>> This is a request for input on what would be a good compromise for general 
>> R users (and not just the R-devel audience).
>> 
>
> Hi,
> Thanks for taking this on - it looks like a big step forward. I have some 
> pretty minor comments/questions. We will, of course, try to adapt the BioC 
> code in time...
>
> We probably want to be sure that any function calling one of these has 
> complete control and can override user defined defaults.

They have arguments to control this, e.g.

> args(chooseCRANmirror)
function (graphics = TRUE)

> I am not completely clear on your model for how the repositories are 
> searched though. Two questions come to mind: 1) How do I get the most recent 
> version of a package, regardless of which repository it is in?

That is the documented default behaviour.  (`Most recent' = `highest 
version number'.)  If more than one repository offers the same version, it 
is taken from the first on the list.

Icens is the only example that I have noticed, where BioC has a later 
version than CRAN.

setRepositories allows users to choose the order only in the text version, 
and you can of course just set the option from the command-line.

> 2) Can a package from one repository have its dependencies resolved in 
> another repository, or not?

Yes, but looking for dependencies is not the default, so you can 
do something like

install.packages("foo", repos="http://www.bioconductor.org",dependencies=TRUE)

to confine the search.

> (this last one is a real can of worms - in my view, as one might really 
> prefer a same repository solution, but that means that repositories 
> might need to completely contain CRAN - which is clearly less 
> desirable). And of course this may be too fine a level of detail, but 
> these problems have come up in our experience.
>
> I think it would be nice if getting dependencies had two (or more 
> levels), so that I could get only those packages that the current 
> package "Depends" on, and a second level where I could get both the 
> "Suggests" and the "Depends". For me (at least) Suggests is weaker, and 
> for many users the "Suggests" set of packages is not always needed - the 
> "Depends" set, is though. Currently the dependencies option seems to 
> only allow TRUE and FALSE.

That's true as for 2.0.0, and we could easily refine it (you need Imports 
too).

Brian

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595