three data.table merge behavior inconsistency

Question

I've been searching around this morning to try to figure out if the failure below is expected but haven't found anything. Could anyone help point me to a related discussion? Otherwise, I might submit as an issue. Appreciate it.

library(data.table)

x <- data.table( a = 1:3 )
y <- data.table( a = 2:4 )
z <- data.table( a = 3:5 )

# works
merge( x , y )
# works
merge( y , z )

# fails
merge( x , merge( y , z ) )
# Error in merge.data.table(x, merge(y, z)) :
#   A non-empty vector of column names for `by` is required.

# works
merge( merge( x , y ) , z )

Ogrady

Looks like this error comes from data.table as it works with data.frame. – Too

Answer 1

This is a clear bug. Please report it. Luckily, it should be easy to fix.

merge.data.table contains this code:

if (is.null(by)) 
  by = intersect(key(x), key(y))
if (is.null(by)) 
  by = key(x)
if (is.null(by)) 
  by = intersect(names(x), names(y))

Now, the issue is that y is keyed (because merge.data.table sets a key):

x <- data.table( a = 1:3 )
y <- merge(data.table( a = 2:4 ), data.table( a = 3:5 ))
haskey(y)
#[1] TRUE

Then,

intersect(key(x), key(y))
#character(0)

Thus, none of the following if conditions is TRUE (we would want the third one to apply here).

This doesn't happen in your last case because of this:

intersect("foo", NULL)
#NULL
intersect(NULL, "foo")
#character(0)

Cirri

thank you! github.com/Rdatatable/data.table/issues/4772 – Ogrady

Answer 2

This was fixed without fanfare in data.table 1.14.4 (May 2022) because the underlying behavior Roland points out of intersect() changed in base R -- intersect(NULL, "foo") is now NULL, which precipitated the fix Roland suggested without realizing the connection to this issue.

merge( x , merge( y , z ) )
# Key: <a>
#        a
#    <int>
# 1:     3

Constituency