zero length .SDcols

[ggrothendieck]

It would be nice if .SDcols could be zero length. In #1786 we discussed this code which works but a special case had to be made for the situation where there were no numeric columns.

iris.dt <- as.data.table(iris)
nums <- which(sapply(iris.dt, is.numeric))
if (length(nums)) iris.dt[, (nums) := .SD/rowMeans(.SD), .SDcols = nums]

It would be nice if .SDcols could allow the specification of no columns so that this would work even in the case that iris has no numeric columns.

iris.dt <- as.data.table(iris)
nums <- which(sapply(iris.dt, is.numeric))
iris.dt[, (nums) := .SD/rowMeans(.SD), .SDcols = nums]

(Incidentally it would also be nice if nums could be logical so that we could have nums <- sapply(iris.dt, is.numeric) . That works in .SDcols= but not on the LHS of := .)

[jangorecki]

Slightly related to #1269.

[HughParsonage]

As a note, there is a (very mild) inconsistency between zero-length character and zero-length integer .SDcols

DT <- data.table(x = 1, y = 2, z = 3)
DT[, .SD, .SDcols = integer(0)]
# Error in `[.data.table`(DT, , .SD, .SDcols = integer(0)) : 
#  .SDcols is numeric but has both +ve and -ve indices
DT[, .SD, .SDcols = character(0)]
# Null data.table (0 rows and 0 cols)

In my view the character(0) response is correct, or at the very least the error message in integer(0) is incorrect.