[Rd] Dispatch mechanism seems to alter object before calling method on it

[Michael Lawrence]

Factors and data.frames are not structures, because they must have a
class attribute. Just call them "objects". They are higher level than
structures, which in practice just shape data without adding a lot of
semantics. Compare getClass("matrix") and getClass("factor").

I agree that inheritance through explicit coercion is confusing. As
far as I know, there are only 2 places where it is used:
1) Objects with attributes but no class, basically "structure" and its
subclasses "array" <- "matrix"
2) Classes that extend a reference type ("environment", "name" and
"externalptr") via hidden delegation (@.xData)

I'm not sure if anyone should be doing #2. For #1, a simple "fix"
would be just to drop inheritance of "structure" from "vector". I
think the intent was to mimic base R behavior, where it will happily
strip (or at least ignore) attributes when passing an array or matrix
to an internal function that expects a vector.

A related problem, which explains why factor and data.frame inherit
from "vector" even though they are objects, is that any S4 object
derived from those needs to be (for pragmatic compatibility reasons)
an integer vector or list, respectively, internally (the virtual
@.Data slot). Separating that from inheritance would probably be
difficult.

Yes, we can consider these to be problems, to some extent stemming
from the behavior and design of R itself, but I'm not sure it's worth
doing anything about them at this point.

Michael

On Wed, May 16, 2018 at 8:33 AM, Hervé Pagès <hpages at fredhutch.org> wrote:
> On 05/15/2018 09:13 PM, Michael Lawrence wrote:
>>
>> My understanding is that array (or any other structure) does not
>> "simply" inherit from vector, because structures are not vectors in
>> the strictest sense. Basically, once a vector gains attributes, it is
>> a structure, not a vector. The methods package accommodates this by
>> defining an "is" relationship between "structure" and "vector" via an
>> "explicit coerce", such that any "structure" passed to a "vector"
>> method is first passed to as.vector(), which strips attributes. This
>> is very much by design.
>
>
> It seems that the problem is really with matrices and arrays, not
> with "structures" in general:
>
>   f <- factor(c("z", "x", "z"), levels=letters)
>   m <- matrix(1:12, ncol=3)
>   df <- data.frame(f=f)
>   x <- structure(1:3, titi="A")
>
> Only the matrix looses its attributes when passed to a "vector"
> method:
>
>   setGeneric("foo", function(x) standardGeneric("foo"))
>   setMethod("foo", "vector", identity)
>
>   foo(f)     # attributes are preserved
>   # [1] z x z
>   # Levels: a b c d e f g h i j k l m n o p q r s t u v w x y z
>
>   foo(m)     # attributes are stripped
>   # [1]  1  2  3  4  5  6  7  8  9 10 11 12
>
>   foo(df)    # attributes are preserved
>   #   f
>   # 1 z
>   # 2 x
>   # 3 z
>
>   foo(x)     # attributes are preserved
>   # [1] 1 2 3
>   # attr(,"titi")
>   # [1] "A"
>
> Also if structures are passed to as.vector() before being passed to
> a "vector" method, shouldn't as.vector() and foo() be equivalent on
> them? For 'f' and 'x' they're not:
>
>   as.vector(f)
>   # [1] "z" "x" "z"
>
>   as.vector(x)
>   # [1] 1 2 3
>
> Finally note that for factors and data frames the "vector" method gets
> selected despite the fact that is( , "vector") is FALSE:
>
>   is(f, "vector")
>   # [1] FALSE
>
>   is(m, "vector")
>   # [1] TRUE
>
>   is(df, "vector")
>   # [1] FALSE
>
>   is(x, "vector")
>   # [1] TRUE
>
> Couldn't we recognize these problems as real, even if they are by
> design? Hopefully we can all agree that:
> - the dispatch mechanism should only dispatch, not alter objects;
> - is() and selectMethod() should not contradict each other.
>
> Thanks,
> H.
>
>>
>> Michael
>>
>>
>> On Tue, May 15, 2018 at 5:25 PM, Hervé Pagès <hpages at fredhutch.org> wrote:
>>>
>>> Hi,
>>>
>>> This was quite unexpected:
>>>
>>>    setGeneric("foo", function(x) standardGeneric("foo"))
>>>
>>>    setMethod("foo", "vector", identity)
>>>
>>>    foo(matrix(1:12, ncol=3))
>>>    # [1]  1  2  3  4  5  6  7  8  9 10 11 12
>>>
>>>    foo(array(1:24, 4:2))
>>>    # [1]  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21
>>> 22 23
>>> 24
>>>
>>> If I define a method for array objects, things work as expected though:
>>>
>>>    setMethod("foo", "array", identity)
>>>
>>>    foo(matrix(1:12, ncol=3))
>>>    #      [,1] [,2] [,3]
>>>    # [1,]    1    5    9
>>>    # [2,]    2    6   10
>>>    # [3,]    3    7   11
>>>    # [4,]    4    8   12
>>>
>>> So, luckily, I have a workaround.
>>>
>>> But shouldn't the dispatch mechanism stay away from the business of
>>> altering objects before passed to it?
>>>
>>> Thanks,
>>> H.
>>>
>>> --
>>> Hervé Pagès
>>>
>>> Program in Computational Biology
>>> Division of Public Health Sciences
>>> Fred Hutchinson Cancer Research Center
>>> 1100 Fairview Ave. N, M1-B514
>>> P.O. Box 19024
>>> Seattle, WA 98109-1024
>>>
>>> E-mail: hpages at fredhutch.org
>>> Phone:  (206) 667-5791
>>> Fax:    (206) 667-1319
>>>
>>> ______________________________________________
>>> R-devel at r-project.org mailing list
>>>
>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__stat.ethz.ch_mailman_listinfo_r-2Ddevel&d=DwIFaQ&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=gynT4YhbmVKZhnX4srXlCWZZRyVBMXG211CKgftdEs0&s=_I0aFHQVnXdBfB5kTLg9TxK_2LHdSuaB6gqZwSx1orQ&e=
>>>
>
> --
> Hervé Pagès
>
> Program in Computational Biology
> Division of Public Health Sciences
> Fred Hutchinson Cancer Research Center
> 1100 Fairview Ave. N, M1-B514
> P.O. Box 19024
> Seattle, WA 98109-1024
>
> E-mail: hpages at fredhutch.org
> Phone:  (206) 667-5791
> Fax:    (206) 667-1319