read_csv fails to guess type of integer column

[tmalsburg]

The data has 73 rows and all entries in the last column are integers. Yet read_csv guesses col_double().

Example:

library(readr)
read_csv("nettle_1999_climate.csv") -> nettle

Results in:

Parsed with column specification:
cols(
  Country = col_character(),
  Population = col_double(),
  Area = col_double(),
  MGS = col_double(),
  Langs = col_double()
)

Last line should be Langs = col_integer().

This is using up-to-date readr from CRAN and R version 3.6.3 (2020-02-29).

[jimhester]

This is actually by design, readr never guesses columns are of type integer.

In earlier versions of readr we did guess columns were of type integer, however this caused numerous issues in practice when the first X number of rows were integers, with floating point numbers later in the file.

Since all 32 bit integers will fit in a 64 bit double this ensures we don't lose information. If you know you want Langs to be integer, you can specify that explicitly, e.g. read_csv("nettle_1999_climate.csv", col_types = list(Langs = "i"))

[tmalsburg]

Makes sense. However, if newer versions of readr don't guess integer, that is important info that should be added to the documentation. Is it okay if I open a documentation issue?

[jimhester]

Sure, it would need to be added at https://github.com/tidyverse/readr/blob/master/vignettes/readr.Rmd#L128.

[tmalsburg]

Shouldn't it go into the man page of read_delim et al? I doubt that users will find it in the vignette.

[gregreich]

Hi, wouldn't it be most consistent if an option like guess_integer would be available for read_csv(), maybe with default value FALSE -- just as there is for parse_guess()? If I know and/or trust my file, I could still benefit from automatic integer detection.

[andysouth]

I came across this because I too didn't know why integers weren't guessed by read_csv() and I wanted them to be :-)

My use case is that I'm reading in tables with ID columns from a standard format (OMOP) that other code expects to be integer and it causes later join failures if they are double.

Found this solution on stackoverflow that seems to work for me.
It reads all columns as character and uses type_convert (also from readr) that does have a guess_integer arg.
read_csv(col_types = cols(.default = "c")) %>% type_convert(guess_integer = TRUE)

(Although I do feel a little nervous deliberately circumventing a design feature ...)

A little more explicit documentation in the man page of read_csv would be useful :-)
Thanks all.