R: Should keys behave this way in data.table?

Question

I have encountered a somewhat unintuitive behavior of keys in data.table package. Here goes an example:

library(data.table)
foo <- data.table(a = c(1:4), b = c(2:5), c = c(3:6), d = c(4:7))
setkey(foo, b)

Then, there is one alarming result of key():

key(foo[, .(mean(c + d)), by = .(b)]) # result is "b".
key(foo[, .(mean(c + d)), by = .(a)]) # result is "a". (!!)

Then, there is another example which produces diffirent, more reasonable results.

foo <- data.table(a = c(4:1), b = c(2:5), c = c(3:6), d = c(4:7))
setkey(foo, b)
key(foo[, .(mean(c + d)), by = .(b)]) # result is "b".
key(foo[, .(mean(c + d)), by = .(a)]) # result is NULL

I admit I'm confused. My lead is this key() somehow checks whether the resulting table needed to be sorted by the elements in by and then assumes it was keyed. Is it a feature? Is it a bug?

Answer 1

Is it a feature? Is it a bug?

In the first example you get key="a" because the result from that query happened to be ordered in a way that a column was in non-decreasing order. Because of that we could call this behaviour a feature.
The problem is that creating a key silently might have not always been desired, thus this behaviour has been changed since you asked that question.
Now (as of 1.12.0) running code from first chunk removes the key and ignores the fact that results are ordered by a.

library(data.table)
foo <- data.table(a = c(1:4), b = c(2:5), c = c(3:6), d = c(4:7))
setkey(foo, b)
key(foo[, .(mean(c + d)), by = .(b)])
#[1] "b"
key(foo[, .(mean(c + d)), by = .(a)])
#NULL