Have finally managed to come up with a fix after checking out sys.calls() from within the as.Date.IDate debugger, which shows something like:
[[1]] rbind(DF, DF) [[2]] rbind(deparse.level, ...) [[3]] `[<-`(`*tmp*`, ri, value = 18042L) [[4]] `[<-.Date`(`*tmp*`, ri, value = 18042L) [[5]] as.Date(value) [[6]] as.Date.IDate(value) I'm not sure why [<- is called, I guess the implementation is to assign to the output block by block? Anyway, we didn't have a [<- method. And [<-.Date looks like: value <- unclass(as.Date(value)) # <- converts to double .Date(NextMethod(.Generic), oldClass(x)) # <- restores 'IDate' class So we can fix our bug by defining a [<- class; the question that I still don't see answered in documentation or source code is, why/where is [<- called, exactly? Mike C On Sun, May 26, 2019 at 1:16 PM Michael Chirico <michaelchiri...@gmail.com> wrote: > Debugging this issue: > > https://github.com/Rdatatable/data.table/issues/2008 > > We have custom class 'IDate' which inherits from 'Date' (it just forces > integer storage for efficiency, hence, I). > > The concatenation done by rbind, however, breaks this and returns a double: > > library(data.table) > DF = data.frame(date = as.IDate(Sys.Date())) > storage.mode(rbind(DF, DF)$date) > # [1] "double" > > This is specific to base::rbind (data.table's rbind returns an integer as > expected); in ?rbind we see: > > The method dispatching is not done via UseMethod(), but by C-internal > dispatching. Therefore there is no need for, e.g., rbind.default. > The dispatch algorithm is described in the source file > (‘.../src/main/bind.c’) as > 1. For each argument we get the list of possible class memberships from > the class attribute. > 2. *We inspect each class in turn to see if there is an applicable > method.* > 3. If we find an applicable method we make sure that it is identical to > any method determined for prior arguments. If it is identical, we proceed, > otherwise we immediately drop through to the default code. > > It's not clear what #2 means -- an applicable method *for what*? Glancing > at the source code would suggest it's looking for rbind.IDate: > > https://github.com/wch/r-source/blob/trunk/src/main/bind.c#L1051-L1063 > > const char *generic = ((PRIMVAL(op) == 1) ? "cbind" : "rbind"); // should > be rbind here > const char *s = translateChar(STRING_ELT(classlist, i)); // iterating over > the classes, should get to IDate first > sprintf(buf, "%s.%s", generic, s); // should be rbind.IDate > > but adding this method (or even exporting it) is no help [ simply defining > rbind.IDate = function(...) as.IDate(NextMethod()) ] > > Lastly, it appears that as.Date.IDate is called, which is causing the type > conversion: > > debug(data.table:::as.Date.IDate) > rbind(DF, DF) # launches debugger > x > # [1] "2019-05-26" <-- singleton, so apparently applied to DF$date, not > c(DF$date, DF$date) > undebug(data.table:::as.Date.IDate) > > I can't really wrap my head around why as.Date is being called here, and > even allowing that, why the end result is still the original class [ > class(rbind(DF, DF)$date) == c('IDate', 'Date') ] > > So, I'm beginning to think this might be a bug. Am I missing something? > [[alternative HTML version deleted]] ______________________________________________ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
