[Rd] Suggestion: Custom filename patterns for non-Sweave vignettes

[Duncan Murdoch]

On 13-02-15 1:53 PM, Henrik Bengtsson wrote:
> Hi Duncan,
>
> thanks you for your prompt reply.
>
>
> On Fri, Feb 15, 2013 at 1:15 AM, Duncan Murdoch
> <murdoch.duncan at gmail.com> wrote:
>> There are several reasons I decided against that:
>>
>>    - two packages may request overlapping patterns, making it much messier to
>> do the matching, checking etc, since the matching would have to depend on
>> the package being processed.
>
> So, isn't that somewhat already taken care of by the 'VignetteBuilder'
> field in DESCRIPTION?  It specifies additional builders in addition to
> the default/builtin Sweave builder.

No, it specifies additional packages besides utils.  Packages may 
specify multiple engines.  For example, knitr can handle Sweave-like 
knitr vignettes, and markdown-based vignettes.  Yihui chose to use the 
same engine for both, but it might make more sense to specify different 
engines.

So a user might say they want a knitr vignette and a .html.rsp vignette. 
  But perhaps in the meantime, Yihui added an engine that can handle 
.rsp files.  So the user would have to list both packages, and there 
would be an ambiguity as to which one should be run.  You might say 
that's the user's problem, but they wouldn't complain to themselves, 
they'd complain to me, so it's my problem.

It would be possible to design all of this to work:  the engine could 
check the file and say "oops, that's not my kind of .rsp file, try the 
next engine".  I just don't think it's worth it.  I certainly don't have 
time to design and program it or even to check your offered patch before 
feature freeze.  I can make small tweaks, but big changes that need lots 
of testing aren't going to happen.


  Conflicts would only happen if a
> package developer (e.g. PkgA) includes a pattern that either (A)
> overrides the builtin in "[.][RrSs](nw|tex)$" / "[.]Rmd$" patterns, or
> (B) specifies to builders with the same patterns.  First of all, there
> are not that many builder packages, so this is something that could be
> negotiated among those to minimize conflicts.  Second, case (A) can be
> protected against by not allowing builder packages (e.g. knitr, rsp,
> ...) to add/register those patterns (tricky but possible to test for)

I don't think it's feasible to check for overlap in regular expression 
patterns.

> (but only default to them if that is what they wish to use).  For case
> (B), the developer of package PkgA has the power to avoid conflicts.
> One could also imagine the ordering of packages listed in
> 'VignetteBuilder' would provide a prioritization.

Sure, but it would be confusing to get an error from knitr when you 
didn't know knitr was handling .rsp.

> BTW, case (A) is basically what the new design is already providing;
> all builder packages use the same patterns.
>
> So, from a package building point of view, I don't see how this would
> make it messier.  I can see that when checking a package it is harder
> to validate matches between input and output formats (is that done?).
> Let me know if I simplifying things too much - then I'll read up on
> the 'R CMD *' source code.
>
>>
>>    - one package may request a pattern that another package uses for
>> auxiliary files, e.g. .bib.  If a user has both types of vignette it would
>> just be a mess.
>
> I see your concern, but is there really a significant risk for this?

If you look through CRAN, you'll see packages that do very weird things. 
  If it's legal, someone will try it.


> And if it would occur, (i) it would be contained to PkgA, (ii) the
> developer of package PkgA would quickly detect it, and (iii) the
> "badly behaving" builder package would rather soon flagged as doing
> something bad (and its developer would be informed and so on).
>
>>
>>    - the extension is also used to determine the output format.  We only
>> support LaTeX (which will be converted to PDF) and HTML output.  It would be
>> reasonable to support direct PDF output, but I don't think any other output
>> formats should be supported.
>
> Yes, supporting PDF output makes sense.  One may also consider
> generation of plain *.txt files (think README.txt and similar).  As I
> see it, the restriction on supported *output* formats are given by
> what the R help system wish to support (which is basically *.pdf and
> *.html documents).  It's clear that the decision on what to support is
> up to the maintainer of the R system (i.e. R core).
>
> When it comes to input/source files for generating those output files,
> it's harder to argue for restrictions.  As I understand it, the new
> support for non-Sweave vignettes is moving away from such restriction,
> which is great.  Despite the restrictions on file extension, it is
> possible to "hijack" (my words) any of the supported extension for
> whatever reason you want, as long as you produce a *.pdf or *.html
> document in the end.  More below...

The issue is that the supplier of a custom input extension would also 
need to specify what kind of output it produced, so R knows how to 
handle it.  That makes it more complicated, harder to test, etc.
>
>>
>> I understand that forcing you to use .Rmd instead of .html.rsp may look
>> unsightly, but I think the extensions need to be fixed, not customizable.
>
> I still find it unfortunate that the R system opens up for processing
> any type of input files but enforces those to have certain filename
> extensions.
>
> As a real example, today Sweave and knitr both use *.Rnw.  This means
> that if I send someone a standalone *.Rnw file, they will not be able
> to tell how to compile it without further instructions from me or by
> inspecting the content type, or by trial and error.  I believe that
> makes reproducible research a bit more tedious.  With unique filename
> extensions, life is easier.  It's easy to imagine that if other
> builder packages (e.g. R.rsp, brew, ...) also start using *.Rnw,
> things are not going to become better.  The current "rules" are
> pushing things in that direction.  To take an extreme stand, it's a
> little bit like using *.txt for all your C, C++, Erlang, Fortran,
> Simula, ... code, because it in the end of the day they all compile to
> binaries anyway.

I agree to some extent, but if sending someone an Rnw causes problems, 
then don't do that.  Rename it before you send it.  Or rename it to Rnw 
when you put it in the vignettes directory.

>
> One may argue that the Rnw/Rtex/Rmd extensions only apply to the R
> package vignettes and you can still use other extensions when you work
> with standalone vignette source files.  That's of course also
> unfortunate, because that will add additional confusion, e.g "You can
> find the vignette in my package, but by the way you should really
> rename it because ...".  The exact same source file will have
> different extensions depending on context.  (In my own case, I found
> *.tex.rsp, *.html.rsp, *.md.rsp, *.Rnw.rsp, ... to be much less
> ambiguous and I prefer not to introduce ambiguity in mapping those to
> *.Rnw/*.Rtex/*.Rmd.)

You could map them to .tex.Rnw, .html.Rmd, .Rnw.Rnw, and your engine 
could do what it does now with *.rsp files.
>
> Finally, the supported extensions are basically *.Rnw, *.Rtex and
> *.Rmd.  To break those down, "*nw" originates from 'Noweb'
> [http://wikipedia.org/wiki/Noweb], "*tex" from TeX
> [http://wikipedia.org/wiki/latex] and "*md" from Markdown
> [wikipedia.org/wiki/Markdown].  The "R*" part indicates that there is
> some additional markup format to those file formats.  But in the end
> of the day, they indicate that the source files should be
> markup-embedded files containing some flavor of Noweb, TeX or
> Markdown.  I find it weird to use those also for, say, formats such as
> HTML, reStructuredText, AsciiDoc, MediaWiki, Org-Mode etc.

That's the etymology, not the meaning.
>
>
> To summarize, I really appreciate the move to a built-in support for
> non-Sweave vignettes (without using custom Makefiles), but I find that
> the supported filename extensions has not been brought along in this
> move.
>

There's always time to argue for R 3.1.0.

Duncan Murdoch