[Rd] package development

Fri Dec 12 15:12:23 CET 2008

Duncan Murdoch wrote:
> On 11/12/2008 6:04 PM, Terry Therneau wrote:
>>   I'm making the move of the survival package from my own environment to,
>> and have stumbled into a vacuum.   The R Extensions manual has really 
>> nice
>> instructions about how to lay out the directories, order the files, and
>> run tests for DISTRIBUTION of a product, but I can't find anything on how
>> to set up a reasonable DEVELOPMENT environment.
>>    In my local world, I had the .c and .s files in a common directory, 
>> with
>> a Makefile that I had created, and the test suite in a subdirectory.  
>> Debugging
>> and development was quite nice.
>>     make
>>     cd test
>>     R
>>     attach("..")
>>         try something and perhaps it fails
>>     q()
>>     cd ..
>> Fix and repeat. The Makefile took some time to create, but paid for 
>> itself a
>> hundred times over.
>>   So, I've now rearranged everything into standard R order.  Then I 
>> did the
>> only thing I could find
>>     R CMD INSTALL ~/myRlib  survival  where "survival" is said 
>> directory.   This turns out to be not useful at all.
>> The survival package is large, and I rather suspected that I would 
>> goof something up, and I did, resulting in the following message
>>
>>     Error in parse(n = -1, file = file) : unexpected end of input at
>>         14450: }
>>         14451:
>> It is not exactly obvious which of the 132 files in my R/ directory is 
>> the culprit here.
> 
> That's something I would like to fix too.  There are (at least) two 
> possible ways:  stop concatenating the files (which is not really needed 
> nowadays, most packages install saved images), or add some markup to the 
> concatenated file so that the parser can report on the original filename 
> and line number (like the #LINE directives output by the C preprocessor).

I would certainly appreciate a fix for this.  When this has happened to 
me, I usually end up just sourcing the individual files into an R 
session until I find the bad file, and get a report with the right line 
number. It always seems like a lot of work for something that should be 
trivial.

> 
>>
>>    In general:
>> 1. The library is large, and recompiling/reparsing everything is very 
>> far from
>> instantaneous.  It is not the edit/load cycle I desire.
> 
> If you install from the directory, the compiling should only be done 
> once (unless you change a file, of course).  (The alternative is 
> installing from the tarball, which is recommended for later stages of 
> testing before distribution, because it's possible something could go 
> wrong in building the tarball.  But it won't include any object files, 
> so you'll recompile every time.)
> 
> You can also use option "--docs=none" to skip building the help system; 
> this will save a bit of time.
> 
> 
>>
>> 2. I take testing seriously: the test suite takes on the order of 15 
>> minutes to
>> run on a fast machine.  I most certainly don't want to run it in the 
>> mid cycle.
> 
> I don't quite follow this.  If you want to run all your tests, you would 
> use R CMD check.  If you only want to run some of them, then you can 
> source things out of the tests directory while running interactively.
> 
>>    Someone must have tackled this.  I'm hoping that there is some 
>> documentation
>> that I have managed to overlook which discussess a good setup for this 
>> middle
>> ground between concieving of a library and packaging it for delivery; the
>> "build, test, make sure it acually works" part of development.

In my experience there are two differ sorts of problems that I think 
could benefit from some improvement. The first is that there should be a 
standard way to have extra tests, that do not get run in the normal CRAN 
testing cycle, or by developers when using a "quick" R CMD check, but 
can be run with a standard mechanism. I do this by putting the tests in 
a separate package, and I have seen reports of different mechanisms, but 
I think they are all somewhat ad hoc. Currently, if you put too much 
testing in your package then all the testing gets omitted on some CRAN 
testing platforms. Just a common directory like extraTests/ would be a 
good start.

The second problem is that a developer usually needs to run tests/ when 
code in R/ has been changed, but probably not run tests/ when the 
changes are only in man/, demos/, or inst/doc/.  The checking that needs 
to be done in those cases is reduced from the full R CMD check.  A 
common way to attack this is with a good Makefile. I wrote an article in 
R News a few years ago about my attempts to do this, and my Makefiles 
are available, but there is some customization necessary, and there is 
lots of room for improvement. It does (mostly) work with make -j, so 
days of testing on a single processor machine can be accomplished in a 
few hours on multicore machines (for me, mileage may vary). I have not 
addressed the idea of trying to specialize files in tests/ to specific 
code files in R/. (I think others have tried to do this with a "unit 
testing" approach.)

Paul Gilbert

> I find the process I follow is to organize the files in the distribution 
> structure from the beginning.   When adding new functions, I'll 
> generally use source() a few times to get the syntax right, and perhaps 
> run simple tests.  (But remember, if you use a NAMESPACE, the functions 
> may not behave the same when they're sourced into the global 
> environment.)  In the early stages, I'll do a lot of installs of the 
> packages.
> 
> If I was porting a big package and wanted to find syntax errors, to work 
> around the not-very-helpful error message you saw I'd do something like
> 
> for (f in list.files("pkg/R", full=TRUE)) source(f)
> 
> which will report the error more informatively.
> 
> Duncan Murdoch
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
====================================================================================

La version française suit le texte anglais.

------------------------------------------------------------------------------------

This email may contain privileged and/or confidential in...{{dropped:26}}