BUG: Improve error message when casting ExtensionDtype to datetime

[arw2019]

We support casting int and float dtypes to DatetimeTZDtype:

In [11]: s = pd.Series(np.random.randint(10, size=3), dtype="int64") 
    ...: s.astype(pd.DatetimeTZDtype(tz="US/Pacific"))                                                  
Out[11]: 
0   1969-12-31 16:00:00.000000004-08:00
1   1969-12-31 16:00:00.000000007-08:00
2   1969-12-31 16:00:00.000000001-08:00
dtype: datetime64[ns, US/Pacific]

We should be able to do this with the corresponding extension dtypes (Int/Float). Currently this errors:

In [12]: s = pd.Series(np.random.randint(10, size=5), dtype="Int64") 
    ...: s.astype(pd.DatetimeTZDtype(tz="US/Pacific"))                                                  
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-12-469a4b536444> in <module>
      1 s = pd.Series(np.random.randint(10, size=5), dtype="Int64")
----> 2 s.astype(pd.DatetimeTZDtype(tz="US/Pacific"))

/workspaces/pandas-arw2019/pandas/core/generic.py in astype(self, dtype, copy, errors)
   5797         else:
   5798             # else, only a single dtype is given
-> 5799             new_data = self._mgr.astype(dtype=dtype, copy=copy, errors=errors)
   5800             return self._constructor(new_data).__finalize__(self, method="astype")
   5801 

/workspaces/pandas-arw2019/pandas/core/internals/managers.py in astype(self, dtype, copy, errors)
    620         self, dtype, copy: bool = False, errors: str = "raise"
    621     ) -> "BlockManager":
--> 622         return self.apply("astype", dtype=dtype, copy=copy, errors=errors)
    623 
    624     def convert(

/workspaces/pandas-arw2019/pandas/core/internals/managers.py in apply(self, f, align_keys, ignore_failures, **kwargs)
    422                     applied = b.apply(f, **kwargs)
    423                 else:
--> 424                     applied = getattr(b, f)(**kwargs)
    425             except (TypeError, NotImplementedError):
    426                 if not ignore_failures:

/workspaces/pandas-arw2019/pandas/core/internals/blocks.py in astype(self, dtype, copy, errors)
    608         if self.is_extension:
    609             try:
--> 610                 values = self.values.astype(dtype)
    611             except (ValueError, TypeError):
    612                 if errors == "ignore":

/workspaces/pandas-arw2019/pandas/core/arrays/integer.py in astype(self, dtype, copy)
    471             na_value = lib.no_default
    472 
--> 473         return self.to_numpy(dtype=dtype, na_value=na_value, copy=False)
    474 
    475     def _values_for_argsort(self) -> np.ndarray:

/workspaces/pandas-arw2019/pandas/core/arrays/masked.py in to_numpy(self, dtype, copy, na_value)
    222             data[self._mask] = na_value
    223         else:
--> 224             data = self._data.astype(dtype, copy=copy)
    225         return data
    226 

TypeError: data type not understood

We should get this to work for both ExtensionArray and the masked versions.

[jbrockmendel]

Does the desirability of this change if #37179 changes how the non-EA case works?

[arw2019]

Does the desirability of this change if #37179 changes how the non-EA case works?

Right. As of #37179 this shouldn't be supported.

Minor but once #37179 is in I'd be keen to align the error message between EA/non-EA cases. Instead of

TypeError: data type not understood    # current EA error message

I'd like to see

TypeError: dtype Int64 cannot be converted to datetime64[ns]  # desired EA error message

[mroeschke]

This looks to work on master now (without the error which makes sense to me). Could use a test

In [1]: In [12]: s = pd.Series(np.random.randint(10, size=5), dtype="Int64")
   ...:     ...: s.astype(pd.DatetimeTZDtype(tz="US/Pacific"))
Out[1]:
0   1969-12-31 16:00:00.000000006-08:00
1   1969-12-31 16:00:00.000000006-08:00
2   1969-12-31 16:00:00.000000002-08:00
3   1969-12-31 16:00:00.000000009-08:00
4   1969-12-31 16:00:00.000000002-08:00
dtype: datetime64[ns, US/Pacific]