BUG: Aggregating over an integer on an empty DataFrame causes datetime64 types to convert to float64

[chrisaycock]

Possibly related to issue #7574:

Suppose I have a DataFrame:

In [61]: df
Out[61]:
          ref                  localtime  size
0    45361866 2014-06-25 14:11:11.753597   100
1    45361866 2014-06-25 14:11:11.753769   100
2    45361866 2014-06-25 14:11:11.754350   200
3    45361866 2014-06-25 14:11:11.756413   200
4    45361866 2014-06-25 14:18:59.442972   200
..        ...                        ...   ...
29  204286294 2014-06-25 19:43:27.770083   100
30  204286294 2014-06-25 19:43:27.771266  7036
31  216308339 2014-06-25 19:57:32.468547   100
32  216308339 2014-06-25 19:57:37.534973   200
33  216308339 2014-06-25 19:57:37.535035   300

[34 rows x 3 columns]

If I aggregate over an integer, the resulting types are as expected:

In [62]: df.groupby('ref').first().reset_index().dtypes
Out[62]:
ref                   int64
localtime    datetime64[ns]
size                  int64
dtype: object

However, if I aggregate on an empty DataFrame, then my timestamp's type is converted from datetime64 to float64:

In [63]: df.query('size>10000').groupby('ref').first().reset_index().dtypes
Out[63]:
index          int64
localtime    float64
size         float64
dtype: object

[jreback]

can you put the frame creation at the top of your example as well. easiest to diagnose when can simply copy-paste

[chrisaycock]

I'm not sure of a clean way to generate a DataFrame with timestamps, but this works in reproducing the error:

df = pd.DataFrame({'ref':[1, 1, 2, 2], 'localtime':4*['2014-06-25 14:11:11.753597'], 'size':[100, 200, 300, 400]})
df.localtime = pd.to_datetime(df.localtime)

[jreback]

works ok in master/0.14.1 (releasing soon). fixed here: #7593

In [1]: df = pd.DataFrame({'ref':[1, 1, 2, 2], 'localtime':4*['2014-06-25 14:11:11.753597'], 'size':[100, 200, 300, 400]})

In [2]: df.localtime = pd.to_datetime(df.localtime)

In [3]: df
Out[3]: 
                   localtime  ref  size
0 2014-06-25 14:11:11.753597    1   100
1 2014-06-25 14:11:11.753597    1   200
2 2014-06-25 14:11:11.753597    2   300
3 2014-06-25 14:11:11.753597    2   400

In [4]: df.dtypes
Out[4]: 
localtime    datetime64[ns]
ref                   int64
size                  int64
dtype: object

In [5]: df.query('size>10000').groupby('ref').first().reset_index().dtypes
Out[5]: 
ref                   int64
localtime    datetime64[ns]
size                  int64
dtype: object

In [6]: df.query('size>10000').groupby('ref').first().reset_index()
Out[6]: 
Empty DataFrame
Columns: [ref, localtime, size]
Index: []

[chrisaycock]

Ah, and it kept your index as ref, which would fix issue #7574. If you can confirm that 0.14.1 passes my test over there, then feel free to close that issue too. Thanks!