Allow map_* to treat NULLs as NAs

[jeremystan]

When working with sparse nested lists (like JSON), it is common to have missing keys or NULL values, which are difficult to coerce into a desired type with purrr.

For example:

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

We can achieve the desired effect by mapping twice and using the null-replace operator

> list(list("a" = 1L), list("b" = 2L)) %>% map("a") %>% map_int(`%||%`, NA)
#> [1]  1 NA

Ideally, map_* and flatten_* would treat NULLs as NAs, possibly by default or if not then by an argument. It could throw a warning when this occurs.

The options I see are:

1. Throw an error
2. Return a vector that excludes the NULL entries
3. Return a vector with appropriate NAs where NULLs were encountered

Currently, purrr implements option 1, which will lead to difficult to debug failures when working with JSON, and a lot of duplicated code to prevent it. I believe option 2 is very dangerous (this is what unlist does). Hence the suggestion for option 3.

[jennybc]

This ground has been explored some before (#110 Default value to replace NULL), which lead to the .null argument.

Does this help?

library(purrr)
list(list("a" = 1L), list("b" = 2L)) %>% map_int("a", .null = NA_integer_)
#> [1]  1 NA

[jeremystan]

I totally missed that (and this issue) - my apologies.

I missed it in part because it's not in the parameters section of the documentation. Should it be added there? It is mentioned under the .f description of the arguments section (which I obviously didn't read closely enough).

[jeremystan]

OK, so on further inspection, the .null argument solves for the missing key, but doesn't solve for the null value:

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

Reading through #110, it seems like the discussion lost the need to treat NULL values as well?

[jennybc]

Good point about explicit NULLs! Obviously that's what I was struggling with too. I don't know if that case was intentionally excluded from the .null treatment? But maybe we will find out ...

[lionel-]

I also think actual NULLs should be replaced by .null. Else we should have called that argument something like .void.