set function crashes the R session

[clerousset]

A very bad usage of set leads to the systematic crash of the Rsession.

# minimal reproductible example :

library(data.table)
n<-1e6
dt<-
  data.table(
    fact=factor(rep("A",n)),
    case=rep(TRUE,n),
    coeff=rep(5,n))
x<-"fact"
 set(
    dt,
    i= which(dt$case==TRUE),
    j= "fact",
    value=dt[(case),..x]*dt$coeff
    )

For the crash to happen :

• n should be big (n=1e4 : no crash, n=1e5 : crash)
• dt$fact needs to be a factor
• argument i needs to be used in set
• argument j needs to be exactly "fact" (needs to overwrite column "fact")
• one should use ..x and not "fact" in dt[(case),..x]

# Output of sessionInfo()

sessionInfo()
R version 4.0.0 (2020-04-24)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 18362)Matrix products: defaultlocale:
[1] LC_COLLATE=French_France.1252 LC_CTYPE=French_France.1252
[3] LC_MONETARY=French_France.1252 LC_NUMERIC=C
[5] LC_TIME=French_France.1252attached base packages:
[1] stats graphics grDevices utils datasets methods baseother attached packages:
[1] data.table_1.13.0loaded via a namespace (and not attached):
[1] compiler_4.0.0 tools_4.0.0

[jangorecki]

Thank you for reporting, are you able to reproduce it on recent version?
On Linux and recent devel I am not having a crash, yet result is corrupted object, that errors when trying to print, so needs some fix anyway.
Worth to note that set is meant to be low overhead function, that can be easily looped million+ times, so probably proper checking for validity of value is not possible.

library(data.table)
n<-1e6
dt<-
  data.table(
    fact=factor(rep("A",n)),
    case=rep(TRUE,n),
    coeff=rep(5,n))
x<-"fact"
 set(
    dt,
    i= which(dt$case==TRUE),
    j= "fact",
    value=dt[(case),..x]*dt$coeff
    )
#Warning message:
#In Ops.factor(left, right) : '*' not meaningful for factors
dt
#Error in as.character.factor(x) : malformed factor

[shrektan]

I can reproduce the crash on Windows 10 with the dev version of data.table. A simpler repro could be this:

library(data.table)
n <- 1e6
dt <- data.table(fact = factor(rep('a', n)))
x <- "fact"
set(dt,
    i = seq_len(1e5),
    j = "fact",
    value = rep(NA, 1e5))

The root cause, IMHO, is that the logical value NA was assigned to an integer vector (factor vectors are in integer), which should be coerced to the numeric NA_INTEGER first.

[jangorecki]

Thanks for simplifying example. Factor is int not double.

str(unclass(factor("a")))
# int 1
# - attr(*, "levels")= chr "a"

So I think it will need to be coerce to NA_integer_.