pivot_longer changes unpivoted date-integer to date-double

[jonspring]

I learned today that dates in R can sometimes be represented as integers, not just as doubles. This can occur like in the example below with seq, which changes the date-double into a date-integer. It seems like an unorthodox data type and might be thought of as a bug of seq that it came to be at all. Unexpectedly, pivot_longer converts these date-integers into the typical date-double, and I couldn't figure out a way to undo that.

I'm not sure where this would cause problems for someone, but I was surprised to see one of my answers on SO did not exactly match the original data because of this subtle change. https://stackoverflow.com/questions/71972396/how-to-reduce-processing-time-of-a-code-in-r

set.seed(42)
df <- data.frame(date_dbl = as.Date( "2022-04-23"),  
                 date_int = seq(as.Date("2022-04-23"), length.out=2, by=1),
                 a = 1:2,
                 b = 3:4)
df
typeof(df$date_dbl)
# [1] "double"
typeof(df$date_int)
# [1] "integer"

df_piv <- tidyr::pivot_longer(df, -(1:2))
typeof(df_piv$date_dbl)
# [1] "double"
typeof(df_piv$date_int)
# [1] "double"

[DavisVaughan]

This is actually intentional. Date vectors are really meant to be double (for better or worse), and it has been a longstanding bug that they are sometimes returned as integer by base R functions. seq.Date() was actually fixed in R 4.2, and now returns a double vector:

x <- seq(as.Date("2022-04-23"), length.out=2, by=1)

typeof(x)
#> [1] "double"

Somewhere under the hood, vctrs is "normalizing" or "fixing" the integer date vector and turning it into a double vector. We generally want this behavior, so I would call this expected.

Unfortunately there isn't a great place to document this, but I expect it to come up less often now that R 4.2+ will generate integer Date vectors less often