[R-pkg-devel] slightly polemic question re R CMD check

R encourages the use of ... particularly in S3 generics, to avoid
over-depending on inheritance to enable flexible use of these generics.
That is, when you call the generic without knowing which class you are
giving it, you cannot specify class-specific arguments. However, some
methods have obvious alternative class-specific behaviors that are
typically enabled using class-specific arguments (e.g. plot). You
cannot support both fully-generic calls and class-specific calls
without giving the generic some flexibility that won't get used in some
cases.


On March 8, 2020 9:41:51 AM PDT, David Hugh-Jones
<davidhughjones at gmail.com> wrote:

Hi Jeff,

I wouldn't say R encourages that in general. Non-generic functions

will

throw an error if you use a non-existent argument. And some generic
functions check for it:

seq(1, 3, blah = 1)
[1] 1 2 3
Warning message:
In seq.default(1, 3, blah = 1) :
extra argument ?blah? will be disregarded

In fact there is even a `chkDots()` function to help with this -

which,

despite having used R or 17 years, I first discovered today :-). So,

it

seems the R base developers thought lenient argument checking could be
a
bad thing, presumably because it lets errors go undetected.

Maybe chkDots is a reasonable workaround. But I wonder what the
rationale
is for R CMD check enforcing that methods *must* be as lenient as the
generic. It seems to lead to a lot of documentation of the form:

@param ... Not used.

Cheers,
David


On Sun, 8 Mar 2020 at 16:24, Jeff Newmiller <jdnewmil at dcn.davis.ca.us>
wrote:

You seem to think this is a bad thing. R does encourage lenient

argument

checking... what rock have you been under for the last 20 years?

On March 8, 2020 5:41:51 AM PDT, David Hugh-Jones <
davidhughjones at gmail.com> wrote:

You're quite right :-) But I think the polemic still holds, because

I

have
to add manual argument checking to all my methods, which has a cost

in

lines of code. Indeed, few base R methods have chosen to do this.

In

effect, the current setup encourages writing methods with "lenient"
argument specifications.

Thank you for the suggestion about ellipsis.

On Sun, 8 Mar 2020, 11:04 G?bor Cs?rdi, <csardi.gabor at gmail.com>

wrote:

You can add the ... argument to chop.default(), and then check

that

length(list(...)) is zero.

Also, you might be interested in the ellipsis package.

Gabor

On Sun, Mar 8, 2020 at 10:56 AM David Hugh-Jones
<davidhughjones at gmail.com> wrote:

Hi all,

My package defines the following method and generic:

chop <- function (x, ...) UseMethod("chop")

chop.default <- function (x, breaks, labels, extend = NULL,

drop

=

TRUE)

{

... }

R CMD check then gives a warning:

W  checking S3 generic/method consistency (695ms)
   chop:
     function(x, ...)
   chop.default:
     function(x, breaks, labels, extend, drop)

   See section ?Generic functions and methods? in the ?Writing

R

   Extensions? manual.

I can fix this by adding a ... argument to chop.default:

chop.default <- function (x, breaks, labels, extend = NULL,

drop

=

TRUE, ...)

But that makes the code less robust because e.g.

chop(x, Breaks = 1:3)

will no longer throw an error from the misspelled argument.

Or I can write:

chop(x, breaks, labels, extend, drop) UseMethod("chop")

but this means I cannot use a different interface for a

different

method.

This seems like a mistake. (That's the polemic.) Or am I

missing

a

better

way? (That's the question.)

David

        [[alternative HTML version deleted]]

      [[alternative HTML version deleted]]

--
Sent from my phone. Please excuse my brevity.