BUG: Resample on empty Period-Indexed Series returns Datetime-Indexed Series

[jduncavage]

Code Sample, a copy-pastable example if possible

In [7]: pd.Series([], pd.period_range(start='2000', periods=0)).resample('W').mean().index Out[7]: DatetimeIndex([], dtype='datetime64[ns]', freq='W-SUN')

Expected Output

PeriodIndex([], dtype='int64', freq='W-SUN')

On pandas: 0.18.0

[max-sixty]

@jreback Lots of issues with PeriodIndex & resample. ref: #12774 & #12770

I didn't really follow the resample refactor so I don't know the genesis of these. I had a look through the current code but can't find easy fixes. Looks like a lot is reliant on the kind attribute, but it's not completely consistent (i.e. a resampled PeriodIndex at times changes kind from period to timestamp, etc).

Do you have any ideas for an overarching strategy for these?
For example, should we just convert every PeriodIndex to a DatetimeIndex, resample it, and then convert back?

[jreback]

@MaximilianR I commented on this in #12774 . The logic from < 0.18.0 is exactly the same. The problem is the period is not systematically tested and there are quite a few edge cases. Further compounding this is the kind kw, which I think we could just totally eliminate. Now the obstacle is that its often easier to change the DTI to a PI, resample then change it back. The code paths are there, just the kind thing really perverts the logic.

So what we need is a fair amount of systematic testing. Then remove the kind (well you don't actually remove it, but deprecate it and we'll remove it later).

[jreback]

I'll make a master Period Resample issue I think.

[max-sixty]

Now the obstacle is that its often easier to change the DTI to a PI, resample then change it back.

OK - can I confirm that you think this is the best path forward, for all PeriodIndex resampling?

The logic from < 0.18.0 is exactly the same.

OK, these are regressions though (from our unit tests failing on our upgrade)

[jreback]

if u have failing tests then they must be things that are not tested in pandas - nothing regressed AFAICT