[Rd] vctrs: a type system for the tidyverse

[Hadley Wickham]

> First off, you are using the word "type" throughout this email; You seem to
> mean class (judging by your Date and factor examples, and the fact you
> mention S3 dispatch) as opposed to type in the sense of what is returned by
> R's  typeof() function. I think it would be clearer if you called it class
> throughout unless that isn't actually what you mean (in which case I would
> have other questions...)

I used "type" to hand wave away the precise definition - it's not S3
class or base type (i.e. typeof()) but some hybrid of the two. I do
want to emphasise that it's a type system, not a oo system, in that
coercions are not defined by superclass/subclass relationships.

> More thoughts inline.
>
> On Mon, Aug 6, 2018 at 9:21 AM, Hadley Wickham <h.wickham using gmail.com> wrote:
>>
>> Hi all,
>>
>> I wanted to share with you an experimental package that I’m currently
>> working on: vctrs, <https://github.com/r-lib/vctrs>. The motivation for
>> vctrs is to think deeply about the output “type” of functions like
>> `c()`, `ifelse()`, and `rbind()`, with an eye to implementing one
>> strategy throughout the tidyverse (i.e. all the functions listed at
>> <https://github.com/r-lib/vctrs#tidyverse-functions>). Because this is
>> going to be a big change, I thought it would be very useful to get
>> comments from a wide audience, so I’m reaching out to R-devel to get
>> your thoughts.
>>
>> There is quite a lot already in the readme
>> (<https://github.com/r-lib/vctrs#vctrs>), so here I’ll try to motivate
>> vctrs as succinctly as possible by comparing `base::c()` to its
>> equivalent `vctrs::vec_c()`. I think the drawbacks of `c()` are well
>> known, but to refresh your memory, I’ve highlighted a few at
>> <https://github.com/r-lib/vctrs#compared-to-base-r>. I think they arise
>> because of two main challenges: `c()` has to both combine vectors *and*
>> strip attributes, and it only dispatches on the first argument.
>>
>> The design of vctrs is largely driven by a pair of principles:
>>
>> -   The type of `vec_c(x, y)` should be the same as `vec_c(y, x)`
>>
>> -   The type of `vec_c(x, vec_c(y, z))` should be the same as
>>     `vec_c(vec_c(x, y), z)`
>>
>> i.e. the type should be associative and commutative. I think these are
>> good principles because they makes types simpler to understand and to
>> implement.
>>
>> Method dispatch for `vec_c()` is quite simple because associativity and
>> commutativity mean that we can determine the output type only by
>> considering a pair of inputs at a time. To this end, vctrs provides
>> `vec_type2()` which takes two inputs and returns their common type
>> (represented as zero length vector):
>>
>>     str(vec_type2(integer(), double()))
>>     #>  num(0)
>>
>>     str(vec_type2(factor("a"), factor("b")))
>>     #>  Factor w/ 2 levels "a","b":
>
>
> What is the reasoning behind taking the union of the levels here? I'm not
> sure that is actually the behavior I would want if I have a vector of
> factors and I try to append some new data to it. I might want/ expect to
> retain the existing levels and get either NAs or an error if the new data
> has (present) levels not in the first data. The behavior as above doesn't
> seem in-line with what I understand the purpose of factors to be (explicit
> restriction of possible values).

Originally (like a week ago 😀), we threw an error if the factors
didn't have the same level, and provided an optional coercion to
character. I decided that while correct (the factor levels are a
parameter of the type, and hence factors with different levels aren't
comparable), that this fights too much against how people actually use
factors in practice. It also seems like base R is moving more in this
direction, i.e. in 3.4 factor("a") == factor("b") is an error, whereas
in R 3.5 it returns FALSE.

I'm not wedded to the current approach, but it feels like the same
principle should apply in comparisons like x == y (even though == is
outside the scope of vctrs, ideally the underlying principles would be
robust enough to suggest what should happen).

> I guess what I'm saying is that while I agree associativity is good for most
> things, it doesn't seem like the right behavior to me in the case of
> factors.

I think associativity is such a strong and useful principle that it
may be worth making some sacrifices for factors. That said, my claim
of associativity is only on the type, not the values of the type:
vec_c(fa, fb) and vec_c(fb, fa) both return factors, but the levels
are in different orders.

> Also, while we're on factors, what does
>
> vec_type2(factor("a"), "a")
>
> return, character or factor with levels "a"?

Character. Coercing to character would potentially lose too much
information. I think you could argue that this could be an error, but
again I feel like this would make the type system a little too strict
and cause extra friction for most uses.

>>     # NB: not all types have a common/unifying type
>>     str(vec_type2(Sys.Date(), factor("a")))
>>     #> Error: No common type for date and factor
>
>
> Why is this not a list? Do you have the additional restraint that vec_type2
> must return the class of one of its operands? If so, what is the
> justification of that? Are you not counting list as a "type of vector"?

You can always request a list, with `vec_type2(Sys.Date(),
factor("a"), .type = list())` - generally the philosophy is too not
make major changes to the type without explicit user input.

I can't currently fully articulate my reasoning for why some coercions
happen automatically, and why some don't. I think these decisions have
to be made somewhat on the basis of pragmatics, and what R users are
currently familiar with. You can see a visual summary of implicit
casts (arrows) + explicit casts (circles) at
https://github.com/r-lib/vctrs/blob/master/man/figures/combined.png.
This matrix must be symmetric, and I think it should be block
diagonal, but I don't otherwise know what the constraints are.

>> On top of this foundation, vctrs expands in a few different ways:
>>
>> -   To consider the “type” of a data frame, and what the common type of
>>     two data frames should be. This leads to a natural implementation of
>>     `vec_rbind()` which includes all columns that appear in any input.
>
>
> I must admit I'm a bit surprised here. rbind is one of the few places that
> immediately come to mind where R takes a fail early and loud approach to
> likely errors (as opposed to the more permissive do soemthing  that could be
> what they meant appraoch of, e.g., out-of-bounds indexing). Are we sure we
> want rbind to get less strict with respect to compatibility of the
> data.frames being combined?

Pragmatically, it's clearly needed for data analysis.

Also note that there are some inputs to rbind that lead to silent data loss:

rbind(data.frame(x = 1:3), c(1, 1000000))
#>   x
#> 1 1
#> 2 2
#> 3 3
#> 4 1

So while it's pretty good in general, there are still a few
infelicities (In particular, I suspect R-core might be interested in
fixing this one)

> Another "permissive" option would be to return a
> data.frame which has only the intersection of the columns. There are
> certainly times when that is what I want (rather than columns with tons of
> NAs in them) and it would be convenient not to need to do the column
> subsetting myself. This behavior would also meet your design goals of
> associativity and commutivity.

Yes, I think that would make sense as an option and would be trivial
to implemet (issue at https://github.com/r-lib/vctrs/issues/46).

Another thing I need to implement is the ability to specify the types
of some columns. Currently it's all or nothing:

vec_rbind(
  data.frame(x = F, y = 1),
  data.frame(x = 1L, y = 2),
  .type = data.frame(x = logical())
)

#>       x
#> 1 FALSE
#> 2  TRUE
#> Warning messages:
#> 1: Lossy conversion from data.frame to data.frame
#> Dropped variables: y
#> 2: Lossy conversion from data.frame to data.frame
#> Dropped variables: y

> I want to be clear, I think what you describe is a useful operation, if it
> is what is intended, but perhaps a different name rather than calling it
> rbind? maybe vec_rcbind to indicate that both rows and columns are being
> potentially added to any given individual input.

Sorry, I should have mentioned that this is unlikely to be the final
name. As well as the problem you mention, I think calling them
vec_cbind() and vec_rbind() over-emphasises the symmetry between the
two operations. cbind() and rbind() are symmetric for matrices, but
for data frames, rbind() is more about common types, and cbind() is
more about common shapes.

Thanks for your feedback, it's very useful!

Hadley

-- 
http://hadley.nz