Reverse dependency tidytable fails due to type mismatch in a merge operation #7604

Activity

1. aitap

   added

   revdepReverse dependenciesReverse dependencies

   on Jan 19, 2026
2. aitap commented on Jan 19, 2026

   

   aitap

   on Jan 19, 2026

   MemberAuthor

   More actions

   This works:

   library(data.table)
   y <- data.table(as.double(2:3), list('foo','bar'))
   x <- data.table(1:3)
   y[x, on='V1']

   But this doesn't:

   (y <- tidytable(x = as.double(2:3), y = list('foo','bar')))
   # # A tidytable: 2 × 2
   #       x y
   #   <dbl> <list>
   # 1     2 <chr [1]>
   # 2     3 <chr [1]>
   (x <- tidytable(x = 1:3))
   # # A tidytable: 3 × 1
   #       x
   #   <int>
   # 1     1
   # 2     2
   # 3     3
   y[x, on = 'x']
   # Error in bmerge(i, x, leftcols, rightcols, roll, rollends, nomatch, mult,  :
   #   typeof x.x (double) != typeof i.x (integer)

   In the second example, x and y are not completely valid (missing the self-reference attribute, not growable). Installing the development version of tidytable fixes the problem, probably due to markfairbanks/tidytable#840.
3. aitap commented on Jan 19, 2026

   

   aitap

   on Jan 19, 2026

   MemberAuthor

   More actions

   The problem is due to bmerge → coerce_col → set changing the data.table in the call frame for coerce_col but not bmerge:
   

   data.table/R/bmerge.R

   Line 25 in 35dbf06

   set(dt, j=col, value=cast_with_attrs(dt[[col]], cast_fun))

   Should bmerge add a check for selfrefok and setalloccol before calling coerce_col?
4. aitap commented on Jan 19, 2026

   

   aitap

   on Jan 19, 2026

   MemberAuthor

   More actions

   A more systematic check for internal use of set() reveals that almost all calls are with a properly intialised data.table, usually created a few lines above, or at least something that is not shared with the caller. The only suspect uses are in setdroplevels
   

   data.table/R/fdroplevels.R

   Line 26 in 35dbf06

   set(x, i=NULL, j=nx, value=fdroplevels(x[[nx]], exclude=exclude))

   
   and bmerge:
   

   data.table/R/bmerge.R

   Line 25 in 35dbf06

   set(dt, j=col, value=cast_with_attrs(dt[[col]], cast_fun))

   Indeed, setdroplevels silently doesn't work on invalid data.tables:

   x <- structure(list(factor('a', levels = letters)), class = c('data.table', 'data.frame'), names = 'x')
   setdroplevels(x)
   levels(x$x)
   #  [1] "a" "b" "c" "d" "e" "f" "g" "h" "i" "j" "k" "l" "m" "n" "o" "p" "q" "r" "s"
   # [20] "t" "u" "v" "w" "x" "y" "z"

   Should we:

   • make the behaviour of set() closer to how it used to work, only requiring setalloccol when changing the number of columns, or
   • change setdroplevels and coerce_col to propagate the re-created data.table to the caller?
5. aitap

   added this to the 1.18.2 milestone on Jan 20, 2026
6. aitap

   closed this as completedin #7606on Jan 20, 2026