date_format does not work correctly on hms objects

[trittweiler]

> library(hms)
> library(scales)
> t <- as.hms(Sys.time())
> date_format("%H:%M", tz="GMT")(t)
[1] "12:30:20.18367"
> format(t, "%H:%M", tz="GMT")
[1] "12:30:20.18367"
> strftime(t, "%H:%M", tz="GMT")
[1] "12:30"

As the function is called "date_format" behaving like "format" is perhaps the right behaviour.

Is this perhaps rather an issue with the hms library which should implement some kind of format method?

[karawoo]

If I understand correctly the issue here is that date_format behaves differently for POSIXct and hms objects, and for hms objects it shows seconds even when the format requested is only "%H:%M". A more illustrative example is below:

library(hms)
library(scales)

t <- Sys.time()

date_format("%H:%M", tz = "GMT")(t)
#> [1] "17:22"
date_format("%H:%M", tz = "GMT")(as.hms(t))
#> [1] "17:22:08.155499"

format(t, "%H:%M", tz = "GMT")
#> [1] "17:22"
format(as.hms(t), "%H:%M", tz = "GMT")
#> [1] "17:22:08.155499"

date_format uses format under the hood, so that is why they behave the same. @hadley do you think this behavior should change? Or is this issue more appropriate for the hms package?

[hadley]

We should probably add support for hms.

[dpseidel]

So this is caused because the format.hms method converts hms objects to character, ultimately calling format.Default and ignoring the format argument to date_format (and the underlying methods format.POSIXlt and format.POSIXct)

Assuming we want to fix this in scales without adjusting the format.hms method itself, the simplest way is to call format.Date directly, although I think this is generally considered bad practice. format.Date and strftime (just a wrapper for format.POSIXlt) seem to differ in how the timezone is treated but (with proper consideration of the timezone) either will output the expected result (as shown using strftime in the OP).

This fix of course stops making sense when a user provides an hms object and asks for a format containing dates not just hours minutes and seconds (as is currently the default for date_format). Perhaps it's worth providing hms support in a separate function time_format? Or is it more desirable to generalize date_format with specific handling of hms objects so they print using a format method for date/time objects?

[hadley]

I'd rewrite date_format() along these lines:

date_format <- function(format = "%Y-%m-%d", tz = 'UTC') {
  function(x) {
    if (inherits(x, "POSIXt")) {
      format(x, format, tz = tz)  
    } else if (inherits(x, "Date")) {
      
    } else if (inherits(x, "difftime")) {
      
    } else {
      stop("Unsupported type (needs better error message)")
    }
      
  }
}

Alternatively, it might be better to add a custom time format type — it's a bit odd to use date_format() with a time.