Deprecating `with` argument makes expr column selector less obvious

[renkun-ken]

Glad to see 1.11.0 version is just released. I noticed from NEWS that with= will be slowly deprecated in future versions (as thoroughly discussed in #2655). However, I find it less obvious for the following cases:

library(data.table)
dt <- data.table(x1 = 1:5, x2 = 5:1, y1 = 0:4, y2 = 4:0)
selector <- list(x = c("x1", "x2"), y = c("y1", "y2"))

To use selector$x or selector$y to select columns in dt, with argument is more obvious.

dt[, selector$x, with = FALSE]

But if with= is no longer available, it seems that we can only store selector$x in another symbol that allows direct access to do the same.

selector_x <- selector$x
dt[, ..selector_x]

Another similar case is calling a function to get the column names we need:

f <- function(prefix, n) {
  paste0(prefix, 1:n)
}
dt[, f("x", 2), with = FALSE]

I'm not sure if I'm missing something new in the latest release but these cases can be ugly without with argument. If .. has to be applied to these cases, it seems that all ..symbol is needed to be directed to symbol in the calling scope (#2589) so that something like dt[, ..selector$x] or dt[, ..f("x")] work.

[jangorecki]

Personally I would prefer keep with argument and set its default dynamically to something like

"[.data.table" = function (x, i, j, with=inspectj(j), ...) ...
# or
"[.data.table" = function (x, i, j, with=NA, ...) {...; if (is.na(with)) auto_detect else force_behavior; ...}
# where auto_detect has logic like `inspectj` but does not have to be separate function

Then user can always force with behavior if needed.

[mattdowle]

On first glance, there seems to be two bugs.

 dt[, ..selector$x]   
[1] "x1" "x2"

j= contains a symbol there ($x) so it's reasonable the result is as if with=TRUE.

dt[, ..selector[["x"]]]
[1] "x1" "x2"

That should work as if with=FALSE because j= contains no symbols. But doesn't. Bug 1 is that ..selector is considered a symbol there even though we know it explicitly isn't a column name due to the prefix.

This already works :

> dt[, paste0("x", 1:2)]
      x1    x2
   <int> <int>
1:     1     5
2:     2     4
3:     3     3
4:     4     2
5:     5     1
> 

But this doesn't :

> dt[, f("x", 2)]
[1] "x1" "x2"

I remember some special casing for paste0 and similar in the internals, which isn't clean. There's an argument to all.vars to exclude function symbols, so turning that on should clean it up (bug 2).

Let's try that first anyway, demo the new fixed behaviour, and then see if you're ok with it.

[jangorecki]

.. / with mentioned in https://nathaneastwood.github.io/2019/08/18/no-visible-binding-for-global-variable/
+1 for not deprecating with but changing default into with=guess_with(j)