Update replace_na() to utilize vctrs

[DavisVaughan]

Revival of #1109
Closes #1168
Closes #912 (by mentioning that you should use across() in the NEWS bullet)

This PR updates replace_na() to utilize vctrs through vec_equal_na() and vec_assign(). This has two breaking changes:

• vec_assign() always casts the RHS to the LHS, but previously with [<- the reverse could be true. I think this is a step in the right direction

  • This means you can't do replace = 1.5 on an integer vector
  • More practically, it means for lists you can't do replace = "foo", it must be replace = list("foo")

• vec_equal_na() is now used for detecting missing values. For list-cols, it will only detect NULL as missing, while the current implementation will replace NULL and empty atomics like integer(). Again, I think this is a net positive.

[DavisVaughan]

Caught a few broken revdeps from this. I still think it is worth pushing forward with this change, sending PRs as required, because these look like real coding errors that we could help fix, and 7 PRs / issues isn't that many. Thoughts @hadley?

(It is especially worth moving forward with this if we also upgrade fill() and drop_na(), as it would be weird to upgrade some but not all to vctrs)

Summary:

• 13 revdeps affected
• 8 unique issues. Generally related to incorrectly using either replace_na(<dbl>, replace = <chr>) or the reverse, replace_na(<chr>, replace = <dbl>). This is stricter now through usage of vec_assign() over [<-.
• 2 unknown issues

---

• causact

  • Through DiagrammeR:::generate_dot() calls replace_na(<dbl>, "")
  • Only call replace_na(replace = "") on character columns rich-iannone/DiagrammeR#448

• heuristicsmineR

  • Same as causact
  • Only call replace_na(replace = "") on character columns rich-iannone/DiagrammeR#448

• processanimateR

  • Same as causact
  • Only call replace_na(replace = "") on character columns rich-iannone/DiagrammeR#448

• RavenR

  • Same as causact
  • Only call replace_na(replace = "") on character columns rich-iannone/DiagrammeR#448

• crosstable

  • Calls replace_na(NULL, "NA") and vec_equal_na(NULL) is an error Should vec_equal_na(NULL) return logical(0)? r-lib/vctrs#1494
  • Return logical() from vec_equal_na(NULL) r-lib/vctrs#1497

• DSSAT

  • Calls replace_na(.data$width, "") where width is a double. Looks like an easy fix
  • Convert $width to character before calling replace_na() palderman/DSSAT#35

• REddyProc

  • Segfault? Doesn't look like us.
  • Re-ran locally and the check() passes fine.

• tidygate

  • Calls replace_na(<chr>, 0)
  • Use a character "0" as the replacement value stemangiola/tidygate#15

• tidytidbits

  • Calls replace_na(<dbl>, "missing")
  • Convert var_name column to character before replacing NAs DavisVaughan/tidytidbits#1
  • Had to send an email linking to that PR, as they are not on GitHub

• Tplyr

  • Was doing replace_na(<int>, replace = list(n = 0)) which promoted the int column to a double. Then they tested that that column was a double. With this PR, it now stays an int.
  • Preemptively convert n to double atorus-research/Tplyr#13

• wpa

  • Calls replace_na(<chr>, 0) through iterating over all columns of a df, regardless of col type
  • Fix: only call replace_na(., 0) on numeric columns microsoft/wpa#195

• xml2relational

  • Calls replace_na(<chr>, 0)
  • Replace with "0" not 0, since the columns are always character jsugarelli/xml2relational#2

• cogmapr

  • Calls replace_na(<dbl>, <chr>)
  • See examples of ?cogmapr::RelationshipTest. Should just convert to character first, I think.
  • Preemptively convert double p.value to character DavisVaughan/cogmapr#1
  • Had to send an email as well
  • On CRAN as of 2021-11-20

[hadley]

If you want to do the work on the revdeps, I think it's reasonable to fix this.

Prior to release, we should also spend a little time running down the "failures to check" and see if we can align more closely to CRAN.

[hadley]

Do we still need this? Or can we rely on vec_assign() to apply the recycling rules appropriately?

[DavisVaughan]

I think we still need this, just because of the way replace_na() works.

Otherwise the replace value would have to be either size 1 or the same size as sum(is.na(data)), and that is almost never the case because you don't know how many missing values there are ahead of time.

x <- c(1, 2, NA, 3, NA)

# This works, of course
replace_na(x, 0)
#> [1] 1 2 0 3 0

# This would actually work too if we removed the restriction
# (because there are two missing values)
# (I don't think we want this to work)
replace_na(x, c(0, -1))
#> [1] 1 2 0 3 -1

# This would not work
# (because replace_na() wants to replace with something of length 2)
y <- c(1, 2, 3, 4, 5)
replace_na(x, y)
# Can't recycle `replace` (size 5) to size 2.