melt drops the class of integer64 columns

[richierocks]

To reproduce, try:

DT <- data.table(x = as.integer64(10^(9:18)))
##                       x
##  1:          1000000000
##  2:         10000000000
##  3:        100000000000
##  4:       1000000000000
##  5:      10000000000000
##  6:     100000000000000
##  7:    1000000000000000
##  8:   10000000000000000
##  9:  100000000000000000
## 10: 1000000000000000000

melt(DT, measure.vars = "x")
##     variable         value
##  1:        x 4.940656e-315
##  2:        x 4.940656e-314
##  3:        x 4.940656e-313
##  4:        x 4.940656e-312
##  5:        x 4.940656e-311
##  6:        x 4.940656e-310
##  7:        x 4.940656e-309
##  8:        x 5.431165e-308
##  9:        x 5.620396e-302
## 10:        x 7.832953e-242

integer64 values are stored using numeric values. The numbers shown in the value column after melting are unclass(DT$x).

[arunsrinivasan]

Thanks. melt up until 1.9.4 did not preserve attributes (this is also documented), to be consistent with reshape2. I'll try to commit the newer implementation of melt in 1.9.6 - and that preserves attributes as expected.

If not, it'll most likely be in 1.9.8.

[richierocks]

OK, thanks. I get the same behaviour with Date and POSIXct objects, but not with factors (which appear as characters not integers). It might be worth including tests on these types with your new version of melt.

[arunsrinivasan]

Yes. It was to be consistent with inconsistency of melt from reshpe2. But I won't do that anymore.

[arunsrinivasan]

This commit 07b90de fixes it. melt now retains attributes wherever possible.