[BUG] loc does not preserve datatype with a single element

[glyg]

Opening a new issue as per @Dr-Irv suggestion on #31763
Sorry issue formatting got lost

import pandas as pd
import numpy as np

not_preserved = pd.DataFrame(np.arange(24).reshape((8, 3)), columns=list("ABC"))
not_preserved['F'] = np.random.random(8)

print("Mixed datatype DataFrame")
print(not_preserved.dtypes)

print("Single value loc: ")
print(not_preserved.loc[0, ['A', 'B', 'C']].dtypes)

print("Single value through slice: ")
print(not_preserved.loc[0:0, ['A', 'B', 'C']].dtypes)

print("The columns are cast to float:")
print(not_preserved.loc[0, ['A', 'B', 'C']])

# This prints:
# Mixed datatype DataFrame
# A      int64
# B      int64
# C      int64
# F    float64
# dtype: object
# Single value loc: 
# float64
# Single value through slice: 
# A    int64
# B    int64
# C    int64
# dtype: object

# The columns are cast to float:
# A    0.0  
# B    1.0
# C    2.0

As a test:

# A single cell is fine

assert isinstance(not_preserved.loc[0, "A"], np.int64) # passes

assert not_preserved.loc[0:0, ["A", "B"]].dtypes["A"] is np.dtype('int64') # passes
assert not_preserved.loc[0, ["A", "B"]].dtypes is np.dtype('int64') # AssertionError

This is with pandas 1.0.3 but was already the case in 0.24.1, it seems that what changed was the behavior of a replace method downstream in my code.
Originally posted by @glyg in #31763 (comment)

[phofl]

Hi, could you please provide your expected output?

[glyg]

Hi, I'm editing the issue for clarity, sorry :D

[glyg]

Sorry the issue formatting got shunt by the quote mechanism. I think it's a pure pandas problem, so the full version info might not be necessary

[phofl]

The resulting Series of ```print(not_preserved.loc[0, ['A', 'B', 'C']].dtypes)`` is converted to a compatible dtype for all columns. There is no attempt to cast it back, when only a subset of columns is selected. You can not keep the dtype for all columns, if they are mixed like int, float and object, so I think this is intended?

[glyg]

Ok, so the idea is that the row term of the loc is applied first, and then the second - selecting the A, B, C columns from the resulting series? I had it in reverse, first getting A, B and C, which are int64 columns and then returning the first row as a Series -- Feel free to close the issue, I guess this is enough that it's there if someone have the same (admittedly corner case) problem.

[jreback]

this is a duplicate issue - there are several discussions about this

if mixing float / int we can't easily do much about this as numpy upcasts

add in an object column for example and a row selection will upcast to the compatible (which would be object)

so actually we could change this by returning an object dtype Series if any mixed types - there is a perf concern here though - would have to check and see how big of a deal this is (and it's hard to tell in the real world)

as i said i think there are some open issues about this

[TomAugspurger]

Agreed that we have other issues about this.