BUG: Fix DatetimeIndex.insert() with strings.

[jtratner]

Falls back to object Index instead. (previously wasn't checking for them), but only strings are allowed.

Fixes #5818.

[jtratner]

This fails tests because of restrictions on partial setting from #4940. I'm not actually convinced that the tests are correct (or at the very least the issue needs to be handled elsewhere). I'll circle back to this tomorrow.

This PR makes DatetimeIndex match Int64Index:

In [2]: ind = pd.Index(range(10))

In [3]: ind.insert(3, '5')
Out[3]: Index([0, 1, 2, u'5', 3, 4, 5, 6, 7, 8, 9], dtype='object')

I'm not clear on why partial setting that has incompatible dtype labels is actually wrong, especially because you can currently do it with non-DatetimeIndex:

In [4]: df = pd.DataFrame(range(10))

In [5]: df
Out[5]:
   0
0  0
1  1
2  2
3  3
4  4
5  5
6  6
7  7
8  8
9  9

[10 rows x 1 columns]

In [6]: df.loc['apple', :] = 2

In [7]: df
Out[7]:
       0
0      0
1      1
2      2
3      3
4      4
5      5
6      6
7      7
8      8
9      9
apple  2

[11 rows x 1 columns]

[jtratner]

You could end up with something similar by concatenating DataFrames with different Index dtypes or by doing a join, so not sure it really makes sense to block partial setting via loc/ix.

[jreback]

jtratner#7

[jtratner]

Just to summarize @jreback 's resolution for those who don't want to follow the link:

1. You can't concatenate floats/ints to DatetimeIndex, because internally DI is represented by int64, so it's too likely you'd get something wrong.
2. You are allowed to concatenate strings.
3. Concatenating something like the string '2001-01-01' should probably also result in object Index

[jreback]

@jtratner just need a release note....then go ahead and merge..thxs

[jreback]

@jtratner this looked fine...thought it was merged...

can you rebase and merge it?

[jreback]

@jtratner ?

[jreback]

@jtratner ?