should `type_convert()` erase the 'spec' attribute?

[wibeasley]

When readr::type_convert() is applied to a data.frame, should it update the 'spec' attribute? Currently, it does not, and potentially returns a misleading result.

library(magrittr); library(readr)
d1 <- 
  read_csv(
    readr_example("mtcars.csv"),
    col_types = cols(.default = col_character())
  )
str(d1)

After the initial creation, the data.frame's spec attribute is correct (i.e., all columns are character).

Classes ‘spec_tbl_df’, ‘tbl_df’, ‘tbl’ and 'data.frame':	32 obs. of  11 variables:
 $ mpg : chr  "21" "21" "22.8" "21.4" ...
 ...
 $ carb: chr  "4" "4" "1" "1" ...
 - attr(*, "spec")=
  .. cols(
  ..   .default = col_character(),
  ..   mpg = col_character(),
  ..   cyl = col_character(),
  ...
  ..   carb = col_character()
  .. )

Then type_convert() correctly converts everything to doubles.

d2 <- data %>% 
  type_convert()

Parsed with column specification:
cols(
  mpg = col_double(),
  cyl = col_double(),
  ...
  carb = col_double()
)

But the spec still shows the character vectors.

spec(d2)

cols(
  .default = col_character(),
  mpg = col_character(),
  cyl = col_character(),
  ...
  carb = col_character()
)

BTW, I really like this function. It was nice when I was stacking a bunch of dataframe on top of each other, but some columns in some sub-dataframes were empty to the data types weren't compatible during the stack.

[jimhester]

spec() is really meant to capture the specification of the original import, I don't think we should change it after the fact. I would be more on board with removing the spec attribute entirely after using type_convert().

[wibeasley]

@jimhester, makes sense to me. I changed the issue title to reflect that.

Do you want me to try a PR?

[jimhester]

Sure, you can actually just change the last line to be df[] and it should drop the spec attribute.

[wibeasley]

@jimhester, thanks for the suggestion. I forgot that was even an option. Would you prefer instead attr(df, "spec") <- NULL? (Placed immediately before df is returned to the caller?)

Does df[] drop all attributes, including others that you may want to retain? (Even if they hadn't been programmed yet.) I'm wondering if someone might call readr::type_convert() on datasets created by something like haven, where they do want to keep those attributes.

[jimhester]

Fair questions, I am fine with doing attr(df, "spec") <- NULL as it is more explicit.

[lock]

This old issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with reprex) and link to this issue. https://reprex.tidyverse.org/