issues with invisible functions in j-expression

[eantonya]

dt = data.table(a = 1:5)

# all is well
boxplot(dt$a)

# lots of errors
dt[, boxplot(a)]
#Error in data.table(stats = c(1, 2, 3, 4, 5), n = 5, conf = c(1.58680503822013,  : 
#  Item 4 has no length. Provide at least one item (such as NA, NA_integer_ etc) to be repeated to #match the 5 rows in the longest column. Or, all columns can be 0 length, for insert()ing rows into.
#In addition: Warning message:
#In data.table(stats = c(1, 2, 3, 4, 5), n = 5, conf = c(1.58680503822013,  :
#  Item 3 is of size 2 but maximum size is 5 (recycled leaving remainder of 1 items)

Related to #482. Iirc it used to plot and output all the stuff you get if you do (boxplot(dt$a)), but now it's not working at all.

[franknarf1]

This makes sense to me since the end result of a j expression is coerced to a data.table.

What I find weird is that dt[, hist(a)] from #1471 works without error, even though it's another case of a list that doesn't coerce to data.table nicely. I don't know the internal bit of code that handles this coercion, but it seems to fail in the latter case because as.data.table.histogram is not a thing, while as.data.table.list is (and boxplot just gives a list).

Personally, I would like to see an option to turn off the coercion (with j returned raw or, when by= is present, as a list column, as in dt[, .(.(boxplot(a))), by=rep(1,5)]).

[jangorecki]

@franknarf1 drop=TRUE could be used for that, where drop=FALSE would force to return data.table by adding additional setDT(list(j)) if j is not a data.table. I've proposed it in another issue where I would benefit from drop=FALSE.

[jangorecki]

Ultimate solution would be to detect visibility of results in j and cascade it down. For now the good alternatives are using DT[,as.null(j)] -> nul or I() function instead of as.null. See related issue #4326.