REGR: empty list selection resets index types

[jreback]

regression from 0.14.0 (which return Index, which is also not correct).

iloc DOES work correctly though

In [9]: df = tm.makeTimeDataFrame()

In [10]: df.index
Out[10]: 
<class 'pandas.tseries.index.DatetimeIndex'>
[2000-01-03, ..., 2000-02-11]
Length: 30, Freq: B, Timezone: None

In [11]: df.loc[[]].index
Out[11]: Float64Index([], dtype='float64')

# this is the correct result
In [4]: df.iloc[[]].index
Out[4]: 
<class 'pandas.tseries.index.DatetimeIndex'>
Length: 0, Freq: None, Timezone: None

In [7]: p = tm.makePanel()

In [8]: p.major_axis
Out[8]: 
<class 'pandas.tseries.index.DatetimeIndex'>
[2000-01-03, ..., 2000-02-11]
Length: 30, Freq: B, Timezone: None

In [9]: p.loc[:,[],:].major_axis
Out[9]: Float64Index([], dtype='float64')

[jreback]

cc @immerrr I don't remember changing anything specifically.....any thoughts?

[cpcloud]

I can take a look at this.

[jreback]

@cpcloud gr8!

[immerrr]

Got it running a debugger for a bit: there's a point when ax.reindex(new_labels) is invoked and at that point new_labels has gone all the way of [] -> np.array([], dtype=float) -> Index(np.array([], dtype=float)) -> Float64Index([]). If you skip conversion to ndarray, you get Index([]) which is what I see on v0.14.0:

In [1]: pd.Series(index=pd.date_range('2013', periods=20))
Out[1]: 
2013-07-17   NaN
2013-07-18   NaN
2013-07-19   NaN
2013-07-20   NaN
2013-07-21   NaN
2013-07-22   NaN
2013-07-23   NaN
2013-07-24   NaN
2013-07-25   NaN
2013-07-26   NaN
2013-07-27   NaN
2013-07-28   NaN
2013-07-29   NaN
2013-07-30   NaN
2013-07-31   NaN
2013-08-01   NaN
2013-08-02   NaN
2013-08-03   NaN
2013-08-04   NaN
2013-08-05   NaN
Freq: D, dtype: float64

In [2]: _1.loc[[]]
Out[2]: Series([], dtype: float64)

In [3]: _1.loc[[]].index
Out[3]: Index([], dtype='object')

In [4]: pd.__version__
Out[4]: '0.14.0'

May be Index.reindex should be smarter about preserving type & metadata of the original container. Incidentally, I've recently encountered the same problem when investigating this issue.

[jreback]

@immerrr thanks, yes I think .reindex should try to preserve if at all possible (might need something special for the 0-len case, but prob just a big buggy somewhere)

[immerrr]

All magic happens here, I don't see anything terribly out of line.

[jreback]

so either take and/or reindex_axis are not preserving (dtype, maybe other attributes) on 0-len then.

[jreback]

bet its getting passed correctly in take, but then re-inferred in the constructor which then returns Index or (something else). prob need to explicity pass in dtype in _simple_new

[immerrr]

It's reindex_axis, yes. _asarray_tuplesafe([]) produces array([], dtype=float) which is then passed to Index.reindex, which brings us to the question above.

[jreback]

yep, which then takes the character of the passed in index.

[jreback]

@immerrr doable for 0.15.0?

[immerrr]

Yes, it's the same as #6552.