[BUG] from_durations seem to be overflowing while converting to a string

[galipremsagar]

Describe the bug
from_durations appears to be overflowing while translating the value to a human readable format, i.e., converting to a string column.

Steps/Code to reproduce bug

In [7]: s = cudf.Series([13645765432432, 134736784, 245345345, -223432411, 999992343241, 3634548734, 23234], dtype='timedelta64[ms]')

In [3]: s.astype('str')
Out[3]: 
0    -55566 days 21:10:41.277551616
1         1 days 13:25:36.784000000
2         2 days 20:09:05.345000000
3        -2 days 14:03:52.411000000
4     11573 days 23:39:03.241000000
5        42 days 01:35:48.734000000
6         0 days 00:00:23.234000000
dtype: object

In [9]: s.to_pandas()
Out[9]: 
0   -55567 days +02:49:18.722448384            #pandas too will overflow but to a different value, probably because they support only `ns` resolution.
1            1 days 13:25:36.784000
2            2 days 20:09:05.345000
3          -3 days +09:56:07.589000
4        11573 days 23:39:03.241000
5           42 days 01:35:48.734000
6            0 days 00:00:23.234000
dtype: timedelta64[ns]

In [10]: s.to_pandas().astype('int')
Out[10]: 
0   -4800978641277551616
1        134736784000000
2        245345345000000
3       -223432411000000
4     999992343241000000
5       3634548734000000
6            23234000000
dtype: int64

In [19]: s.astype('int64')
Out[19]: 
0    13645765432432
1         134736784
2         245345345
3        -223432411
4      999992343241
5        3634548734
6             23234
dtype: int64


In [12]: import numpy as np

In [14]: def delta_format(delta: np.timedelta64) -> str:      # A method to calculate part of `timedelta` i.e., days & hours.
    ...:     days = delta.astype("timedelta64[D]") / np.timedelta64(1, 'D')
    ...:     hours = int(delta.astype("timedelta64[h]") / np.timedelta64(1, 'h') % 24)
    ...: 
    ...:     if days > 0 and hours > 0:
    ...:         return f"{days:.0f} d, {hours:.0f} h"
    ...:     elif days > 0:
    ...:         return f"{days:.0f} d"
    ...:     else:
    ...:         return f"{hours:.0f} h"
    ...: 


In [20]: s[0]   # The scalar value is correctly preserved.
Out[20]: numpy.timedelta64(13645765432432,'ms')

In [15]: delta_format(s[0])
Out[15]: '157937 d, 2 h'             # Expecting this value while we convert to string's at index s[0]

In [16]: delta_format(s[1])
Out[16]: '1 d, 13 h'

Expected behavior

Does not overflow in strings i.e., return 157937 d, 2hrs.. at index 0 of s.

Environment overview (please complete the following information)

• Environment location: [Bare-metal]
• Method of cuDF install: [from source]

Additional context
This bug surfaced in: #11749

[galipremsagar]

cc: @davidwendt

[davidwendt]

The duration value is eventually passed to the cuda::std::chrono to be dissected into components.
We generally do not do data validation in libcudf. Bad data in gives bad data out.
General discussion comes from #5505

[GregoryKimball]

Is this issue still open? The repro doesn't appear to overflow on 22.12.

>>> s = cudf.Series([13645765432432, 134736784, 245345345, -223432411, 999992343241, 3634548734, 23234], dtype='timedelta64[ms]')
>>> s.astype('str')
0    157937 days 02:23:52.432
1         1 days 13:25:36.784
2         2 days 20:09:05.345
3        -2 days 14:03:52.411
4     11573 days 23:39:03.241
5        42 days 01:35:48.734
6         0 days 00:00:23.234
dtype: object
>>> 

[galipremsagar]

IDK what change fixed this issue in 22.12(works for me too), @davidwendt shall we close this issue or keep it open if you think this isn't actually fixed?

[davidwendt]

My assertion was that there is nothing for us to fix per our developer guidelines https://github.com/rapidsai/cudf/blob/branch-22.12/cpp/doxygen/developer_guide/DEVELOPER_GUIDE.md#libcudf-does-not-introspect-data
So I'm ok to close this.