[C++] strptime rolls-over dates not in range for current month

[asfimport]

I noticed some potentially unexpected behaviour when converting from string to date. Days that are out of bounds for the given month are rolled over into the following month.

I think the expected behaviour would be to either error (Python) or return NULL/NA (R), but not to roll over dates in the following month.

library(arrow, warn.conflicts = FALSE)
library(lubridate, warn.conflicts = FALSE)
library(dplyr, warn.conflicts = FALSE)

df <- tibble::tibble(string_date = "1999-02-30")

# base R returns NA
df %>% 
  mutate(date = strptime(string_date, format = "%Y-%m-%d"))
#> # A tibble: 1 × 2
#>   string_date date  
#>   <chr>       <dttm>
#> 1 1999-02-30  NA

# arrow rolls over the 30th of February into the 2nd of March
df %>% 
  arrow_table() %>% 
  mutate(date = strptime(string_date, format = "%Y-%m-%d")) %>% 
  collect()
#> # A tibble: 1 × 2
#>   string_date date               
#>   <chr>       <dttm>             
#> 1 1999-02-30  1999-03-02 00:00:00

Thanks Alenka, Joris and Rok for helping me with the Python examples:
pandas:

>>> import pandas as pd
>>> pd.to_datetime("1999-02-30", format="%Y-%m-%d")
...
ValueError: time data 1999-02-30 doesn't match format specified

datetime:

>>> import datetime
>>> from datetime import datetime
>>> datetime.strptime("1999-02-30", "%Y-%m-%d")
...
ValueError: day is out of range for month

arrow:

>>> import pyarrow.compute as pc
>>> print(pc.strptime("1999-02-30", format="%Y-%m-%d", unit="s"))
1999-03-02 00:00:00

Reporter: Dragoș Moldovan-Grünfeld / @dragosmg
Watchers: Rok Mihevc / @rok

Related issues:

• [C++] Add error handling option to StrptimeOptions (is related to)

_{Note: This issue was originally created as ARROW-15948. Please see the migration documentation for further details.}

[asfimport]

Joris Van den Bossche / @jorisvandenbossche:
Copying my comment from the other issue: personally, I don't like that behaviour, but I suppose we get this from the system strptime? (so that might even depend on your OS?)
It might be interesting to check what date.h's version of strptime does.

I think the expected behaviour would be to either error (Python) or return NULL/NA (R),

I would say it is both: by default error, and optionally return null (after ARROW-15665)

[asfimport]

Rok Mihevc / @rok:
Two notes:

Perhaps system strptimes have an option to disable this behaviour?

Switching from system to date.h strptime might have performance implications so we should probably benchmark as we change this. 

[asfimport]

Antoine Pitrou / @pitrou:
Does date.h have a strptime? We use a vendored implementation from musl on Windows.

[asfimport]

Rok Mihevc / @rok:
It does, but there's this comment: https://github.com/apache/arrow/blob/master/cpp/src/arrow/util/value_parsing.h#L784

[asfimport]

Antoine Pitrou / @pitrou:
Ah, well, let's not use it, then :)

[asfimport]

Rok Mihevc / @rok:
But maybe it's nice to dates? :D