Weird behaviour using `.SD`, `by=` and `copy`

[jhrmnn]

I've stumbled upon a really weird behaviour with 1.9.4. The minimal example is

library(data.table)
dt <- data.table(a=c(1, 1, 1, 0, 0),
                 b=c("A", "B", "A1", "A", "B"))
dt
##    a  b
## 1: 1  A
## 2: 1  B
## 3: 1 A1
## 4: 0  A
## 5: 0  B
dt[, nrow(.SD[b == 'B']), by=.(a)]
##    a V1
## 1: 1  1
## 2: 0  0
dt[, nrow(copy(.SD)[b == 'B']), by=.(a)]
##    a V1
## 1: 1  1
## 2: 0  1
dt[3, b:="C"]
##    a b
## 1: 1 A
## 2: 1 B
## 3: 1 C
## 4: 0 A
## 5: 0 B
dt[, nrow(.SD[b == 'B']), by=.(a)]
##    a V1
## 1: 1  1
## 2: 0  1
dt[, nrow(copy(.SD)[b == 'B']), by=.(a)]
##    a V1
## 1: 1  1
## 2: 0  1

Note that when dt[3, b] == "A1", the output with and without copy differs, which I don't understand at all. When dt[3, b] == "C", the behaviour is as expected.

[arunsrinivasan]

Thanks, I've reproduced the bug on 1.9.5 as well.

For the future, please read and follow the instructions here and here.

[arunsrinivasan]

The issue is due to automatic indexing creating secondary key attribute on .SD.

require(data.table)
options(datatable.auto.index = FALSE)
dt <- data.table(a=c(1, 1, 1, 0, 0),
                 b=c("A", "B", "A1", "A", "B"))
dt[, nrow(.SD[b == 'B']), by=.(a)]
#    a V1
# 1: 1  1
# 2: 0  1

[jangorecki]

the issue was solved by following line:
is.null(attr(x, '.data.table.locked'))) { # fix for #958, don't create auto index on '.SD'.
isn't it unsafe to use attr(..., exact=FALSE)?
I know it is unlikely but such error caused by some approximation of names can be hard to debug if the errors occurs.