API/BUG/CLN: Index/Series common dtype handling

[sinhrks]

Code Sample, a copy-pastable example if possible

Some inconsistencies there.

# OK: specifying int dtype with float value coerces to int (same as numpy)
pd.Series([1, 2, 3.5], dtype=int)
#0    1
#1    2
#2    3
# dtype: int64

# OK
pd.Index(np.array([1, 2, 3.5], dtype=int))
# Int64Index([1, 2, 3], dtype='int64')

# NG: If list is given, dtype is ignored
pd.Index([1, 2, 3.5], dtype=int)
# Float64Index([1.0, 2.0, 3.5], dtype='float64')

Expected Output

• All list-like input should be handled by a consistent logic.
• If dtype is specified, the result must be the specified dtype. Otherwise raise.
  • To do this, merge base Index creation and Series sanitizing logic to use same ndarray dtype handling.

output of pd.show_versions()

Current master.

[jreback]

cc @gfyoung

this is very much like #15187

this might even be closable now, can you have a look?

[gfyoung]

Similar, but not the same. This relates to how we coerce dtype in general (not the uint64 corner case). However, note that what @sinhrks is suggesting as the correct behavior contradicts what you were proposing in [12] of that issue you opened.

[gfyoung]

I think what might be best is to open one issue that combines all of these dtype casting issues if that works with everyone.

[gfyoung]

@jreback : Opened #15832, which includes this issue. I think we can close this one then.