Allow map_*() to consult .null for length 0 elements

[jennybc]

It's baaack. See #231 and #110. When map_*() is used for indexing, we can now use .null to specify a value if the key is missing or has explicit value of NULL. But now I am struggling with the case where the element is present, is not NULL, but has length zero.

library(purrr)
list(list("a" = 1L), list("a" = list(), "b" = 2L)) %>% 
  map_int("a", .null = NA_integer_)
#> Error: Result 2 is not a length 1 atomic vector

I have real examples of this phenomenon from the GitHub API.

cc @jeremystan

[lionel-]

Should this be another argument .empty?

[jennybc]

At least when JSON is the source of the list, it seems morally ok to handle via .null and not introduce another argument. That's what I took away from this conversation re: jsonlite, which is the source of my list() elements (jeroen/jsonlite#154).

[lionel-]

otoh NULL and empty vectors are really different things in R.

It's true that we already conflate void elements with NULL values, but that's mainly because void elements are generally treated as NULL in R due to its lisp roots.

[jennybc]

I see your point. But once you've committed to making an atomic vector of a certain type, I would guess that void, NULL, and empty almost always become NA. If they need to be handled differently, maybe people should do some work with map() first? I suppose that is what I could / should be doing above.

[hadley]

Related to #199

[lionel-]

This feels wrong to me, conflating the notions of null and empty. Empty is a property of vectors, while null is a property of any R object (which are all option types thanks to null).

We have the same problem in rlang. is_empty() returns FALSE for NULL, but we don't have a predicate that returns TRUE for both empty vectors and NULL. I think that if the notions had to be conflated (which I don't think is a good idea), it should be that empty vectors can also be NULL rather than NULL objects can also be empty vectors, which this change implies.

It would still be useful to have a term that means "either NULL or empty". Maybe lengthless, since NULL has length 0. Or "blank"?

[lionel-]

To be more precise, I don't think it's a bad idea to treat NULLs and empty the same way, I do it all the time by checking the length, e.g. if (!length(x)) .... Just that we should promote a consistent vocabulary for these fundamental R properties.

[hadley]

We were partially doing this anyway and it makes life easier; but maybe we should change .null to .empty?

[lionel-]

yes I now see the rlang definition for is_empty() is in fact based on length > 0. So maybe it wouldn't be too bad to consider that "empty" = empty + null?

[hadley]

Yeah. I sometimes think of NULL as an object is a vector with no type but length 0.

[jennybc]

You could attack this from another angle and name the argument .default, instead of .null or .empty. This conveys it is used whenever the input does not provide anything suitable but doesn't make specific assertions about why the input is unsuitable. There is precedent for default, both with and without the dot, in dplyr::recode(), dplyr::nth(), dplyr::lead(), dplyr::lag(), etc.

[jennybc]

In fact, dplyr seems to have a function for this: dplyr:::default_missing().

[hadley]

Another option would be .absent

[lionel-]

I like .default as well.