BUG: inconsistencies/errors in quantile on empty DataFrame

[jorisvandenbossche]

In PR #14536, I added some tests in comments, as they currently fail or give inconsistent results:

1. Empty frame with float dtype:

df = DataFrame(columns=['a', 'b'], dtype='float64')
df.quantile(0.5)
df.quantile(0.5, axis=1)

In 0.18.1, this gives NaNs or empty frame depending on the axis (which is correct I think):

In [10]: df.quantile(0.5)
Out[10]: 
a   NaN
b   NaN
dtype: float64

In [9]: df.quantile(0.5, axis=1)
Out[9]: 
Empty DataFrame
Columns: []
Index: []

But on master, the axis=1 case errors (df.quantile(0.5) also gives NaNs):
master:

In [8]: df.quantile(0.5, axis=1)
...
ValueError: need at least one array to concatenate

2. Empty frame with int dtype

df = DataFrame(columns=['a', 'b'], dtype='int64')
df.quantile(0.5)

Opposed to float dtype giving a series of NaNs, with integers it gives an empty frame in 0.18.1:

In [11]: df.quantile(0.5)
Out[11]: 
Empty DataFrame
Columns: []
Index: []

and on master also raises the ValueError as for float with axis=1:

In [14]: df.quantile(0.5)
...
ValueError: need at least one array to concatenate

3. Empty frame with datetime values

df = DataFrame(columns=['a', 'b'], dtype='datetime64')
df.quantile(0.5, numeric_only=False)

On 0.18.1 / master, it gives a series of NaNs, where this should be NaTs:

In [13]: df.quantile(0.5, numeric_only=False)
Out[13]: 
a   NaN
b   NaN
Name: 0.5, dtype: float64

4. Frame with only only datetime columns but without only_numeric=False

df = DataFrame({'a': pd.to_datetime(['2010', '2011']), 'b': [0, 5], 'c': pd.to_datetime(['2011', '2012'])})
df[['a', 'c']].quantile(.5)

On 0.18.1, this gives an empty frame

In [8]: df[['a', 'c']].quantile(.5)
Out[8]: 
Empty DataFrame
Columns: []
Index: []

while on master this raises the same ValueError as above:

In [7]: df[['a', 'c']].quantile(.5)
...
ValueError: need at least one array to concatenate

[rogeriomgatto]

There are also inconsistencies in groupby behaviour (0.19.1):

>>> df = pd.DataFrame({'a': np.zeros(0, dtype='int'), 'b': np.zeros(0, dtype='int')})

>>> df
Empty DataFrame
Columns: [a, b]
Index: []

>>> df.dtypes
a    int64
b    int64
dtype: object

>>> df.groupby(level=0).mean()
Empty DataFrame
Columns: [a, b]
Index: []

>>> df.groupby(level=0).quantile(0.5)
Empty DataFrame
Columns: []
Index: []

[jreback]

@rogeriomgatto pull-requests to fix are welcome, and get things done!

[mroeschke]

Looks like all these cases work on master (except the df = DataFrame(columns=['a', 'b'], dtype='datetime64') df.quantile(0.5, numeric_only=False) case returning NaTs). Could use test for the other cases.

We can create an independent issue for the case that is not working once these regression tests are added.