Makes NumericIndex constructor dtype aware

[oguzhanogreden]

• closes Non-unique integers coerced to float during UInt64Index creation with explicit #29526
• enables Correct type inference for UInt64Index during access #29420
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff
• whatsnew entry

[jschendel]

Thanks! Approach looks good. Can you add an entry to the 1.0 whatsnew under the "Bug Fixes - Numeric" subsection (around line 339)?

[jschendel]

I don't think this should rely on dtype being explicitly passed in (it's None by default) so maybe make this dtype=self._dtype for now since the dtype parameter isn't being validated.

[oguzhanogreden]

This initially made sense to me.

However, numpy seems to complicate this suggestion since it'll cast input if dtype is explicitly passed:

In [1]: import numpy as np 
   ...: np.__version__                                                                                                                                                                                                                                                                                                         
Out[1]: '1.17.3'

In [2]: np.array([1, "2"]) 
   ...:                                                                                                                                                                                                                                                                                                                        
Out[2]: array(['1', '2'], dtype='<U21')

In [3]: np.array(["1", "2"])  
   ...:                                                                                                                                                                                                                                                                                                                        
Out[3]: array(['1', '2'], dtype='<U1')

In [4]: np.array(["1", "2"], dtype='int')                                                                                                                                                                                                                                                                                      
Out[4]: array([1, 2])

This test suggests that pandas keeps the same behavior. With 458c25e, I tried to check the type of elements. The side effect of that is consuming iterators, leading to the current errors.

[oguzhanogreden]

The numpy behavior I report above is brought up here but not acted on.

[oguzhanogreden]

I added the what's new entry but I must admit I haven't yet managed to set up building docs in a performant way. I'll check that soon and let you know.

[oguzhanogreden]

The failing tast was this. Now there are more of them :) Perhaps it's too late now, I'll take a fresh look over the weekend.

I had some questions as well...

2. Index._coerce_to_ndarray() has practically become Index._check_types_and_coerce_to_ndarray(). Does it make sense to make an issue/PR to split these?
3. The change I made is in the method of an abstract Index class. This method has effectively become a "numeric-only" method. However, CategoricalIndex also inherits from Index. Does it make sense to change the conditional added in 458c25e to reflect this, something like

# use _typ instead of inferred_type property since this method, if ever used, will be used before the instance exists
if cls._typ != 'categoricalindex' and lib.infer_dtype(data) == 'string':

[pep8speaks]

Hello @oguzhanogreden! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2019-11-17 21:30:54 UTC

[oguzhanogreden]

Sorry about a bit of back and forth here.

Finally:

1. The solution depends on explicit passing of dtype.
2. @jschendel 's point about argument validation seems to be addressed centrally here.

[jreback]

lgtm. some small requests. ping on green.

[jreback]

hah, this is what we had quite some time ago. thanks

[jreback]

thanks @oguzhanogreden

[jschendel]

Sorry, I've been a little behind on reviewing and noticed a couple small things that could be done as a follow up. Edit: Nevermind, this can be ignored.