BUG: regression in Timedelta.__rfloordiv__ with integer data

[jorisvandenbossche]

From a failing example in the docs (http://pandas-docs.github.io/pandas-docs-travis/timeseries.html#from-timestamps-to-epoch):

In [46]: stamps = pd.date_range('2012-10-08 18:15:05', periods=4, freq='D')

In [47]: stamps.view('int64') // pd.Timedelta(1, unit='s')
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-47-9893b9bae6cf> in <module>()
----> 1 stamps.view('int64') // pd.Timedelta(1, unit='s')

/home/joris/scipy/pandas/pandas/_libs/tslibs/timedeltas.pyx in pandas._libs.tslibs.timedeltas.Timedelta.__rfloordiv__()

TypeError: Invalid dtype int64 for __floordiv__

This was working correctly in 0.22, and is mentioned in the docs explicitly as a way to get unix epoch data.

The previously working example (http://pandas.pydata.org/pandas-docs/stable/timeseries.html#from-timestamps-to-epoch) looks like:

In [42]: stamps.view('int64') // pd.Timedelta(1, unit='s')
Out[42]: array([1349720105, 1349806505, 1349892905, 1349979305])

cc @jbrockmendel

[jorisvandenbossche]

I think this was introduced (on purpose, tests were added) in #18961.
I agree the above example that was included in the docs is a bit strange, as when dividing 'int / timedelta', the timedelta is implicitly cast to an integer as it nanoseconds. But, on the other hand, this was clearly in our docs as a way to get epoch data, so we cannot just break it I think

[jorisvandenbossche]

So we need to make a decision on this one. Options:

1. Keep the behaviour on master (= keep limitation of behaviour of Timedelta.__rfloordiv__ ) and update the docs with an alternative solution. What would this alternative be?

2. Undo the regression and re-enable the previous behaviour of Timedelta.__rfloordiv__

3. Re-enable the previous behaviour of Timedelta.__rfloordiv__ but with a deprecation warning it will be removed in the future. Provide alternative solution in the warning and docs.

cc @jreback @TomAugspurger

Given that the method is not very intuitive, I would be in favor of option 3 if we have a good alternative.

[TomAugspurger]

I think 3.

…

On Thu, Mar 29, 2018 at 4:05 AM, Joris Van den Bossche < ***@***.***> wrote: So we need to make a decision on this one. Options: 1. Keep the behaviour on master (= keep limitation of behaviour of Timedelta.__rfloordiv__ ) and update the docs with an alternative solution. What would this alternative be? 2. Undo the regression and re-enable the previous behaviour of Timedelta.__rfloordiv__ 3. Re-enable the previous behaviour of Timedelta.__rfloordiv__ but with a deprecation warning it will be removed in the future. Provide alternative solution in the warning and docs. cc @jreback <https://github.com/jreback> @TomAugspurger <https://github.com/TomAugspurger> Given that the method is not very intuitive, I would be in favor of option 3 *if* we have a good alternative. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#19761 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIsGUBITOCGfju8lKqvuG1KGNDI5wks5tjKP2gaJpZM4SKMmk> .

[jbrockmendel]

For 1, the alternative solution would be to also use the integer value for the Timedelta

stamps.view('int64') // pd.Timedelta(1, unit='s').value

It's also a bit weird to have __floordiv__ work but not __div__ (as is the case in 0.22.0)

[jreback]

so either use .value on the rhs operand (as @jbrockmendel suggest) or

In [12]: (stamps-pd.Timestamp(0)) // pd.Timedelta(1, 's')
Out[12]: Int64Index([1349720105, 1349806505, 1349892905, 1349979305], dtype='int64')

is a reasonable soln (This is more natural as you are dividing a TDI / Timedelta which is a bit more clear)

[jorisvandenbossche]

Are we comfortable with advertising the use of .value for Timestamp/Timedeltas ?

In that case, the stamps.view('int64') // pd.Timedelta(1, unit='s').value seems maybe best. Although we also had discussion regarding deprecating view (of course we can also use .astype('int') (but that does an extra copy) or .asi8)