Feature Request: pass column to key

[ptoche]

# make a data.table
set.seed(1)
dt <- data.table(id = 1:5, x1 = 1:5, x2 = 5:1, x3 = round(runif(5, 1, 5), 0), key = "id")

I can define the data.table with either id = 1:10 or "id" = 1:10, but I must define the key with key = "id" as key = id does not work:

# Feature Request:  ``key = id``  
dt <- data.table(id = 1:5, x1 = 1:5, x2 = 5:1, x3 = round(runif(5, 1, 5), 0), key = id)
##Error in data.table(id = 1:5, x1 = 1:5, x2 = 5:1, x3 = round(runif(5,  : 
##  object 'id' not found

[jangorecki]

Any proposal how to handle following key definition?

library(data.table)
k = "b"
dt = data.table(b = "a", k = "b", key = k)

Just put your expected results.

[MichaelChirico]

@jangorecki this is in relation to a SO question about consistency as it seems some data.table functions (e.g., subset.data.table) accept expressions while others don't.

The relevant code of data.table is here:

if (!is.null(key)) {
        if (!is.character(key)) 
            stop("key argument of data.table() must be character")
        if (length(key) == 1L) {
            key = strsplit(key, split = ",")[[1L]]
        }
        setkeyv(value, key)
    }

This snippet comes after the assignment of the columns; seems reasonable for there to be a way to switch setkeyv to setkey, perhaps with a with option.

There is where things get dicey, though, because leaving with = TRUE as the default (as in [.data.table]) would break existing code.

Another alternative would be to interpret lists as expressions (as in by/keyby) by default, so that something data.table-y like data.table(b = "a", k = "b", key = .(k)) would set the key to k while data.table(b = "a", k = "b", key = k) would obey the current behavior (i.e. set the key to b).

[franknarf1]

I can define the data.table with either id = 1:10 or "id" = 1:10

This is a basic feature of R and not specific to data.table:

c(a   = 1)   # ok
c("a" = 1)   # ok
c(a   = "b") # ok
c(a   = b)   # nope

something data.table-y like data.table(b = "a", k = "b", key = .(k)) would set the key to k while data.table(b = "a", k = "b", key = k) would obey the current behavior (i.e. set the key to b).

Yeah, I was thinking along those lines too. One problem with that is that it is inconsistent with how .() works in other places, like by=k vs by=.(k), which do the same thing, with the former just being shorthand.

On the other hand by="k,x", by=.(k,x) and by=c("k","x") all work, so I'm sure it can be done to match the OP's expectations. I just dunno how much work it is... there are other places where such syntax (for raw names without quotes) would also be nice, like in .SDcols=: #642 And there's also syntax like .SDcols=V1:V3 (which currently works, even though .SDcols=V1 does not) that might be nice to have in many places. I guess there's an almost endless array of nonstandard-evaluation shortcuts one might ask for...

[MichaelChirico]

@franknarf1 Good point about by=k vs. by=.(k); what comes to mind there is the use of character vectors on the LHS of :=. So perhaps key=k should default to setting column k as the key, while key=(k) sets b?

[jangorecki]

Closing this as won't fix. It doesn't seems to bring anythin more useful than 2 quote character less to type, and it introduces the complexity of catching key argument unevaluated. Moreover the problem described almost 5 years ago my be, about such interface, has not been addressed.