[C++] Provide a way to go from numeric to duration

[asfimport]

Update: I keep forgetting that R integers are translated into int32 (since R does not have int64). int64 -> duration is supported. 
 Would it make sense to support float -> duration or int32 -> duration?

======================
Currently it is not possible to directly create a duration object from a numeric one (for example through casting).

library(arrow)

a <- Array$create(32L)
a$cast(duration("s"))
#> Error: NotImplemented: Unsupported cast from int32 to duration using function cast_duration
#> /Users/dragos/Documents/arrow/cpp/src/arrow/compute/function.cc:231  DispatchBest(&inputs)

This underpins a lot of the date-time arithmetic in R, which support the conversion/ coercion of an integer to difftime (R's equivalent for duration), such as in the pipeline below.

library(arrow, warn.conflicts = FALSE)
#> See arrow_info() for available features
library(dplyr, warn.conflicts = FALSE)
library(lubridate, warn.conflicts = FALSE)

df <- tibble(time = as_datetime(c("2022-03-07 15:00:28", "2022-03-06 14:00:28"))) 
df
#> # A tibble: 2 × 1
#>   time               
#>   <dttm>             
#> 1 2022-03-07 15:00:28
#> 2 2022-03-06 14:00:28

df %>% 
  mutate(time2 = time + seconds(2))
#> # A tibble: 2 × 2
#>   time                time2              
#>   <dttm>              <dttm>             
#> 1 2022-03-07 15:00:28 2022-03-07 15:00:30
#> 2 2022-03-06 14:00:28 2022-03-06 14:00:30

Reporter: Dragoș Moldovan-Grünfeld / @dragosmg

Related issues:

• [C++] Cast to duration from types other than int64 (is duplicated by)
• [R] Helper function for casting from float to duration via int64() (relates to)
• [C++][R] Can't cast integer difftime to duration (but works with double) (supercedes)

_{Note: This issue was originally created as ARROW-15862. Please see the migration documentation for further details.}

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:

Would it make sense to support float -> duration or int32 -> duration?

Do you mean in the (C++) Cast machinery?

We also don't support this for example timestamp:

# casting int64 to timestamp is fine
>>> pa.array([1, 2, 3]).cast(pa.timestamp("s"))
<pyarrow.lib.TimestampArray object at 0x7fb5c314c340>
[
  1970-01-01 00:00:01,
  1970-01-01 00:00:02,
  1970-01-01 00:00:03
]

# casting int32 not
>>> pa.array([1, 2, 3], type="int32").cast(pa.timestamp("s"))
... ArrowNotImplementedError: Unsupported cast from int32 to timestamp using function cast_timestamp
../src/arrow/compute/function.cc:231  DispatchBest(&inputs)

We might want to do this, but I think it is a general issue / question for the casting machinery of Arrow how to handle casts that might need an additional step (like "int32 -> timestamp" can be done as "int32 -> int64 -> timestamp" which are all supported)

[asfimport]

Dragoș Moldovan-Grünfeld / @dragosmg:
Yes, I was thinking in the C++ cast machinery. True, at the moment I am going via int64, which is not a problem. Float -> duration would be useful. R for example can handle conversions of double/ float to difftime (the base R duration equivalent).

as.difftime(32.5, units = "secs")
#> Time difference of 32.5 secs

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
Related issue, but for other kernels: ARROW-11508

[asfimport]

Antoine Pitrou / @pitrou:
@dragosmg You can cast indirectly, for example float->int64->duration. I don't think we want to proliferate variations of cast kernels, though this could be discussed. @lidavidm Thoughts?

[asfimport]

David Li / @lidavidm:
Yeah, this seems like something we could implement in the bindings or frontend by just doing two casts. Unless there's a significant performance benefit to avoiding the extra cast (I would guess not?) we should at least try to be a little conscious about code/binary size.

[asfimport]

Antoine Pitrou / @pitrou:
Also, I think the last cast (int64->duration) is zero-copy already?

[asfimport]

Dragoș Moldovan-Grünfeld / @dragosmg:
@pitrou I don't think float -> int64 casting works. Unless I am missing something.

library(arrow)
  
a <- Array$create(32.5)
a$cast(float())$cast(int64())
#> Error: Invalid: Float value 32.5 was truncated converting to int64

[asfimport]

Antoine Pitrou / @pitrou:
Well, if the cast would truncate, it fails by default but you can select an unsafe cast:

>>> pc.cast(3.5, pa.int64())
Traceback (most recent call last):
  ...
ArrowInvalid: Float value 3.5 was truncated converting to int64
>>> pc.cast(3.5, pa.int64(), safe=False)
<pyarrow.Int64Scalar: 3>

[asfimport]

Dragoș Moldovan-Grünfeld / @dragosmg:
True, but an unsafe cast ultimately alters the value.  

a <- Array$create(32.6)
b <- Array$create(32.7)
a$cast(int64(), safe = FALSE)$cast(duration()) + b$cast(int64(), safe = FALSE)$cast(duration())
#> Array
#> <duration[s]>
#> [
#>   64
#> ]

[asfimport]

Neal Richardson / @nealrichardson:
I'm ok with not supporting float cast to duration, but it seems that any integer type should be able to cast to duration, not just int64. This seems odd to me:

> Scalar$create(1L)$cast(duration())
Error: NotImplemented: Unsupported cast from int32 to duration using function cast_duration

[github-actions]

This issue has been marked as stale because it has had no activity in the past 365 days. Please remove the stale label or comment below, or this issue will be closed in 14 days. If this improvement is still desired but has no current owner, please add the 'Status: needs champion' label.