Not a reproducible example but clearly shows there is something to improve. When ignore.row.order=F then column by column comparison goes using all.equal.numeric method (that has tolerance), when ignore.row.order=T, then AFAIR we roll join on all fields which relies on setNumericRounding() for numeric columns.
sc = \(x) setkey(copy(x))
all.equal(d, b, ignore.row.order = TRUE)
#[1] "Dataset 'current' has rows not present in 'target'"
all.equal(sc(d), sc(b))
#[1] TRUE
1.16.2
Not a reproducible example but clearly shows there is something to improve. When ignore.row.order=F then column by column comparison goes using all.equal.numeric method (that has tolerance), when ignore.row.order=T, then AFAIR we roll join on all fields which relies on setNumericRounding() for numeric columns.
1.16.2