R devel - Mar 2019 - as.Date(Inf) displays as 'NA' but is actually 'Inf'

Hi Gabriel,

The point is that it *visually* displays as NA, but is.na() still 
responds as FALSE.

When I (and I am sure many people) see an NA, we then use is.na(). If we 
see Inf displayed, we then use is.infinite(). With as.Date() this breaks 
down.

I'm not arguing that as.Date(Inf) should be coerced to NA. I'm arguing 
that as.Date(Inf) should be *visually* displayed as Inf (i.e. the 
truth!). I doubt this would break any existing code, because 
as.Date(Inf) acts as Inf in every way possible, except for when you 
visually look at the output printed on the screen.

William - For all the other Date bugs, they don't visually display false 
information about the variable's contents. They might give wrong output, 
but the output displayed is what exists inside the variable.

If we can't trust the R console to display the truth, then we are in a 
lot of trouble.

 > a <- as.Date(Inf, origin="2018-01-01")
 > a
[1] NA
 > is.na(a)
[1] FALSE

Richard

Gabriel Becker wrote on 06/03/2019 00:33:
> Richard,
>
> Well others may chime in here, but from a mathematical point of view, 
> the concept of "infinite days from right now" is well-defined, so
it
> maybe a "valid" date in that sense, but what day and month it
will be
> (year will be Inf) are indeterminate/not well defined. Those are 
> rightfully, NA, it seems?
>
> I mean you could disallow dates to take Inf at all, ever. I don't feel 
> strongly one way or the other about that, personally. That said, if 
> inf dates are allowed, its not clear to me that displaying the 
> "Formatted" date string as NA, even if the value isn't,? is
wrong
> given it can't be determined for that "date" is. It could be
displayed
> differently, I suppose, but all the ones I can think of off the top of 
> my head would be problematic and probably break lots of 
> formatted-dates parsing code out there in the wild (and in R, I would 
> guess). Things like displaying "Inf-NA-NA", or just
"Inf". Neither of
> those are going to handle a read-write round-trip well, I think.
>
> So my personal don't-really-have-a-hat-in-the-ring opinion would be to 
> either leave it as is, or force as.Date(Inf, bla) to actually be NA.
>
> Best,
> ~G
>
> On Tue, Mar 5, 2019 at 12:06 PM Richard White <w at rwhite.no 
> <mailto:w at rwhite.no>> wrote:
>
>     Hi,
>
>     I think I've discovered a bug in base R.
>
>     Basically, when using 'Inf' as as 'Date', is is
visually displayed as
>     'NA', but R still treats it as 'Inf'. So it is very
confusing to work
>     with, and can easily lead to errors:
>
>     # Visually displays as NA
>     ?> as.Date(Inf, origin="2018-01-01")
>     [1] NA
>
>     # Visually displays as NA
>     ?> str(as.Date(Inf, origin="2018-01-01"))
>     Date[1:1], format: NA
>
>     # Is NOT NA
>     ?> is.na <http://is.na>(as.Date(Inf,
origin="2018-01-01"))
>     [1] FALSE
>
>     # Is still Inf
>     ?> is.infinite(as.Date(Inf, origin="2018-01-01"))
>     [1] TRUE
>
>     This gets really problematic when you are collapsing dates over
>     groups
>     and you want to find the first date of a group. Because min() returns
>     Inf if there is no data:
>
>     # Visually displays as NA
>     ?> as.Date(min(), origin="2018-01-01")
>     [1] NA
>     Warning message: In min() : no non-missing arguments to min;
>     returning Inf
>
>     # Visually displays as NA
>     ?> str(as.Date(min(), origin="2018-01-01"))
>     Date[1:1], format: NA
>     Warning message: In min() : no non-missing arguments to min;
>     returning Inf
>
>     # Is not NA
>     ?> is.na <http://is.na>(as.Date(min(),
origin="2018-01-01"))
>     [1] FALSE
>     Warning message: In min() : no non-missing arguments to min;
>     returning Inf
>
>     # This is bad!
>     ?> as.Date(min(), origin="2018-01-01") >
"2018-01-01"
>     [1] TRUE
>     Warning message: In min() : no non-missing arguments to min;
>     returning Inf
>
>     Here is my sessionInfo():
>
>     ?> sessionInfo()
>     R version 3.5.0 (2018-04-23)
>     Platform: x86_64-pc-linux-gnu (64-bit)
>     Running under: Debian GNU/Linux 9 (stretch)
>     Matrix products: default
>     BLAS: /usr/lib/openblas-base/libblas.so.3
>     LAPACK: /usr/lib/libopenblasp-r0.2.19.so
>     <http://libopenblasp-r0.2.19.so>
>
>     locale:
>     [1] LC_CTYPE=C.UTF-8 LC_NUMERIC=C LC_TIME=C.UTF-8 LC_COLLATE=C.UTF-8
>     LC_MONETARY=C.UTF-8
>     [6] LC_MESSAGES=C LC_PAPER=C.UTF-8 LC_NAME=C LC_ADDRESS=C
>     LC_TELEPHONE=C
>     [11] LC_MEASUREMENT=C.UTF-8 LC_IDENTIFICATION=C
>
>     attached base packages:
>     [1] stats graphics grDevices utils datasets methods base loaded via a
>     namespace (and not attached):
>     [1] compiler_3.5.0 tools_3.5.0 yaml_2.1.19
>
>     ?> Sys.getlocale()
>     [1]
>    
"LC_CTYPE=C.UTF-8;LC_NUMERIC=C;LC_TIME=C.UTF-8;LC_COLLATE=C.UTF-8;LC_MONETARY=C.UTF-8;LC_MESSAGES=C;LC_PAPER=C.UTF-8;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=C.UTF-8;LC_IDENTIFICATION=C"
>
>     ______________________________________________
>     R-devel at r-project.org <mailto:R-devel at r-project.org>
mailing list
>     https://stat.ethz.ch/mailman/listinfo/r-devel
>

	[[alternative HTML version deleted]]

On Tue, Mar 5, 2019 at 9:54 PM Richard White <w at rwhite.no> wrote:
> Hi Gabriel,
>
> The point is that it *visually* displays as NA, but is.na() still
> responds as FALSE.
>
> When I (and I am sure many people) see an NA, we then use is.na(). If we
> see Inf displayed, we then use is.infinite(). With as.Date() this breaks
> down.
>
> I'm not arguing that as.Date(Inf) should be coerced to NA. I'm
arguing
> that as.Date(Inf) should be *visually* displayed as Inf (i.e. the truth!).
> I doubt this would break any existing code, because as.Date(Inf) acts as
> Inf in every way possible, except for when you visually look at the output
> printed on the screen.
>
> William - For all the other Date bugs, they don't visually display
false
> information about the variable's contents. They might give wrong
output,
> but the output displayed is what exists inside the variable.
>
> If we can't trust the R console to display the truth, then we are in a
lot
> of trouble.
>
Well, I think it (subtly) actually is the truth though. What is displayed
when you print a date is the *formatted date string, not the numeric value
stored within the date*. The formatted date string of the infinite date, is
actually, correctly,  NA, because, for the reasons I pointed out in my last
post, it is indeterminate.
> x = as.Date(Inf, origin = "2018-01-01")
> format(x)
[1] NA


So that is what is happening, both technically, but also conceptually. For
the record, I'd be surprised by that too, but I think its a situation of
pieces working correctly individually, but together having a correct but
unintuitive behavior.

Others may feel differently though, thats just my read on it.

Best,
~G


> > a <- as.Date(Inf, origin="2018-01-01")
> > a
> [1] NA
> > is.na(a)
> [1] FALSE
>
> Richard
>
> Gabriel Becker wrote on 06/03/2019 00:33:
>
> Richard,
>
> Well others may chime in here, but from a mathematical point of view, the
> concept of "infinite days from right now" is well-defined, so it
maybe a
> "valid" date in that sense, but what day and month it will be
(year will be
> Inf) are indeterminate/not well defined. Those are rightfully, NA, it
> seems?
>
> I mean you could disallow dates to take Inf at all, ever. I don't feel
> strongly one way or the other about that, personally. That said, if inf
> dates are allowed, its not clear to me that displaying the
"Formatted" date
> string as NA, even if the value isn't,  is wrong given it can't be
> determined for that "date" is. It could be displayed differently,
I
> suppose, but all the ones I can think of off the top of my head would be
> problematic and probably break lots of formatted-dates parsing code out
> there in the wild (and in R, I would guess). Things like displaying
> "Inf-NA-NA", or just "Inf". Neither of those are going
to handle a
> read-write round-trip well, I think.
>
> So my personal don't-really-have-a-hat-in-the-ring opinion would be to
> either leave it as is, or force as.Date(Inf, bla) to actually be NA.
>
> Best,
> ~G
>
> On Tue, Mar 5, 2019 at 12:06 PM Richard White <w at rwhite.no> wrote:
>
>> Hi,
>>
>> I think I've discovered a bug in base R.
>>
>> Basically, when using 'Inf' as as 'Date', is is
visually displayed as
>> 'NA', but R still treats it as 'Inf'. So it is very
confusing to work
>> with, and can easily lead to errors:
>>
>> # Visually displays as NA
>>  > as.Date(Inf, origin="2018-01-01")
>> [1] NA
>>
>> # Visually displays as NA
>>  > str(as.Date(Inf, origin="2018-01-01"))
>> Date[1:1], format: NA
>>
>> # Is NOT NA
>>  > is.na(as.Date(Inf, origin="2018-01-01"))
>> [1] FALSE
>>
>> # Is still Inf
>>  > is.infinite(as.Date(Inf, origin="2018-01-01"))
>> [1] TRUE
>>
>> This gets really problematic when you are collapsing dates over groups
>> and you want to find the first date of a group. Because min() returns
>> Inf if there is no data:
>>
>> # Visually displays as NA
>>  > as.Date(min(), origin="2018-01-01")
>> [1] NA
>> Warning message: In min() : no non-missing arguments to min; returning
Inf
>>
>> # Visually displays as NA
>>  > str(as.Date(min(), origin="2018-01-01"))
>> Date[1:1], format: NA
>> Warning message: In min() : no non-missing arguments to min; returning
Inf
>>
>> # Is not NA
>>  > is.na(as.Date(min(), origin="2018-01-01"))
>> [1] FALSE
>> Warning message: In min() : no non-missing arguments to min; returning
Inf
>>
>> # This is bad!
>>  > as.Date(min(), origin="2018-01-01") >
"2018-01-01"
>> [1] TRUE
>> Warning message: In min() : no non-missing arguments to min; returning
Inf
>>
>> Here is my sessionInfo():
>>
>>  > sessionInfo()
>> R version 3.5.0 (2018-04-23)
>> Platform: x86_64-pc-linux-gnu (64-bit)
>> Running under: Debian GNU/Linux 9 (stretch)
>> Matrix products: default
>> BLAS: /usr/lib/openblas-base/libblas.so.3
>> LAPACK: /usr/lib/libopenblasp-r0.2.19.so
>>
>> locale:
>> [1] LC_CTYPE=C.UTF-8 LC_NUMERIC=C LC_TIME=C.UTF-8 LC_COLLATE=C.UTF-8
>> LC_MONETARY=C.UTF-8
>> [6] LC_MESSAGES=C LC_PAPER=C.UTF-8 LC_NAME=C LC_ADDRESS=C
LC_TELEPHONE=C
>> [11] LC_MEASUREMENT=C.UTF-8 LC_IDENTIFICATION=C
>>
>> attached base packages:
>> [1] stats graphics grDevices utils datasets methods base loaded via a
>> namespace (and not attached):
>> [1] compiler_3.5.0 tools_3.5.0 yaml_2.1.19
>>
>>  > Sys.getlocale()
>> [1]
>>
>>
"LC_CTYPE=C.UTF-8;LC_NUMERIC=C;LC_TIME=C.UTF-8;LC_COLLATE=C.UTF-8;LC_MONETARY=C.UTF-8;LC_MESSAGES=C;LC_PAPER=C.UTF-8;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=C.UTF-8;LC_IDENTIFICATION=C"
>>
>> ______________________________________________
>> R-devel at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>
>
>
	[[alternative HTML version deleted]]

>>>>> Gabriel Becker 
>>>>>     on Tue, 5 Mar 2019 22:01:37 -0800 writes:
    > On Tue, Mar 5, 2019 at 9:54 PM Richard White <w at rwhite.no>
wrote:
    >> Hi Gabriel,
    >> 
    >> The point is that it *visually* displays as NA, but is.na() still
    >> responds as FALSE.
    >> 
    >> When I (and I am sure many people) see an NA, we then use is.na().
If we
    >> see Inf displayed, we then use is.infinite(). With as.Date() this
breaks
    >> down.
    >> 
    >> I'm not arguing that as.Date(Inf) should be coerced to NA.
I'm arguing
    >> that as.Date(Inf) should be *visually* displayed as Inf (i.e. the
truth!).
    >> I doubt this would break any existing code, because as.Date(Inf)
acts as
    >> Inf in every way possible, except for when you visually look at the
output
    >> printed on the screen.
    >> 
    >> William - For all the other Date bugs, they don't visually
display false
    >> information about the variable's contents. They might give
wrong output,
    >> but the output displayed is what exists inside the variable.
    >> 
    >> If we can't trust the R console to display the truth, then we
are in a lot
    >> of trouble.
    >> 

    > Well, I think it (subtly) actually is the truth though. What is
displayed
    > when you print a date is the *formatted date string, not the numeric
value
    > stored within the date*. The formatted date string of the infinite
date, is
    > actually, correctly,  NA, because, for the reasons I pointed out in my
last
    > post, it is indeterminate.

    >> x = as.Date(Inf, origin = "2018-01-01")

    >> format(x)

    > [1] NA


    > So that is what is happening, both technically, but also conceptually.
For
    > the record, I'd be surprised by that too, but I think its a
situation of
    > pieces working correctly individually, but together having a correct
but
    > unintuitive behavior.

    > Others may feel differently though, thats just my read on it.

    > Best,
    > ~G

Thank you Richard and Gabe and Bill (Dunlap),
I agree with both of you that the behavior is suprising (to > 99.9% of
useRs).

Gabe very nicely explains how it happens and also why it does
make some sense *and* that a change may be problematic.

However, the "principle of least surprise" I've learned very long
ago
from Doug Bates is good "guiding" principle for software design
(if you allow to weight it with other principles, etc).

Here is a bit of slightly more principled code to show the
phenomenon, including the fact noticed by Bill that both
as.Date() and format.Date() should probably be tweaked such as
to signal warnings (e.g. on integer overflow for too large numbers).

## -------------------------------------------------------------------------
xDates <- lapply(c(-Inf, Inf, NA, NaN,
                   1e9, 4e9, 1e100, .Machine$double.xmax),
                 as.Date, origin = "2000-01-01")
str(xDates) # --> first 4 *all* show as  NA
sapply(xDates, is.na) # the two +-Inf are not NA
(f.D <- sapply(xDates, format))# 1..4: NA, then "negative" but all
the same (?!)
stopifnot(is.na(f.D)[1:4]) # the formats (of 1..4) *are* all NA !!
## show their true internals -- still contain what was put there :
for(d in xDates) dput(d)
## -------------------------------------------------------------------------

produces
> xDates <- lapply(c(-Inf, Inf, NA, NaN,
+                    1e9, 4e9, 1e100, .Machine$double.xmax),
+                  as.Date, origin =
"2000-01-01")
> str(xDates) # --> first 4 *all* show as  NA
List of 8
 $ : Date[1:1], format: NA
 $ : Date[1:1], format: NA
 $ : Date[1:1], format: NA
 $ : Date[1:1], format: NA
 $ : Date[1:1], format: "2739907-01-04"
 $ : Date[1:1], format: "-5877641-06-23"
 $ : Date[1:1], format: "-5877641-06-23"
 $ : Date[1:1], format: "-5877641-06-23"
> sapply(xDates, is.na) # the two +-Inf are not NA
[1] FALSE FALSE  TRUE  TRUE FALSE FALSE FALSE FALSE
> (f.D <- sapply(xDates, format))# 1..4: NA, then "negative" but
all the same (?!)
[1] NA               NA               NA               NA              
"2739907-01-04"  "-5877641-06-23"
[7] "-5877641-06-23"
"-5877641-06-23"
> stopifnot(is.na(f.D)[1:4]) # the formats (of 1..4) *are* all NA !!
> ## show their true internals -- still contain what was put there :
> for(d in xDates) dput(d)
structure(-Inf, class = "Date")
structure(Inf, class = "Date")
structure(NA_real_, class = "Date")
structure(NaN, class = "Date")
structure(1000010957, class = "Date")
structure(4000010957, class = "Date")
structure(1e+100, class = "Date")
structure(1.79769313486232e+308, class =
"Date")
> 
---------

What if we left NA ( NA_character_ specifically ) as result for format(),
but changed the print() method so it gives better information
here ?

I would argue that -Inf and Inf should show differently than
true NA's or NaN's .. not the least because infinitely past and
infinitely into the future are different concepts.

Martin Maechler
ETH Zurich (and R Core team)