Prevent adding new attributes to the accessors .str, .dt and .cat

[jankatins]

assigning to Series.str, Series.dt, or Series.cat was not failing
although you couldn't get the value back:

In[10]: a = pandas.Series(pandas.Categorical(list("abc")))
In[11]: a.cat.labels = [1,2]
In[12]: a.cat.labels
Traceback (most recent call last):
  File "C:\portabel\miniconda\envs\pandas_dev\lib\site-packages\IPython\core\interactiveshell.py", line 2883, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-12-a68ee763e4e8>", line 1, in <module>
    a.cat.labels
AttributeError: 'CategoricalAccessor' object has no attribute 'labels'

Now we fail early:

In[10]: a = pandas.Series(pandas.Categorical(list("abc")))
In[11]: a.cat.labels = [1,2]
Traceback (most recent call last):
  File "C:\data\external\pandas\pandas\tests\test_categorical.py", line 3633, in test_cat_accessor_no_new_attributes
    c.cat.labels = [1,2]
  File "C:\data\external\pandas\pandas\core\base.py", line 121, in __setattr__
    raise AttributeError( "You cannot add any new attribute '{key}'".format(key=key))
AttributeError: You cannot add any new attribute 'labels'

Closes: #10673

[max-sixty]

I've had some pickling issues with doing this in the past, as __setattr__ gets called during unpickling, without __init__ being called.
Another way - slightly less elegant but lower blast radius - is to set a private variable _str and then have a property def str(self): ; return self._str. That'll also prevent setting of str.

[jreback]

I agree here with @MaximilianR
this is already done in NDFrame using _accessors as a white-list with __getattr__, so should be straightforward, so I think would be ok to emulate in PandasDelegate (not that these allow sub-classes to override, e.g. Series does for .dt,.str.,.cat themselves.

[jankatins]

I'm not sure I can follow: as far as I can see, the accessors .str, .dt and .cat are not pickled (and they should not!). This is not trying to prevent s.str = something but s.str.someattr = something. The main benefit is that a user, who misspells an attribute, is not surprised. E.g. when using s.cat.labels instead of s.cat.categories.

Currently assigning (s.cat.labels = "aaa") works, but you can't get the value from this attribute back (print(s.cat.labels) raises an AttributeError). I've actually no idea where this behaviour comes from, the last should work, as the first (without this PR) uses the normal item assignment and the latter normal get (that's at least what the pycharm debugger lets me see).

[jankatins]

This needs a rebase if pulled after #11582...

[jreback]

this inherits from PandasObject so not necessary

[jankatins]

At least during my tests it was? It inherits from PandasDelegate:

class CategoricalIndex(Index, PandasDelegate):

[jankatins]

Ok, took me a while to find a place for the mixin: both core.common and core.base resulted in circular depencies due to the way the StringMethods are defined (str = ...(StringMethods,...) -> can't get the import to a function :-():

Traceback (most recent call last):
  File "C:\Program Files (x86)\JetBrains\PyCharm Community Edition 143.165\helpers\pycharm\utrunner.py", line 121, in <module>
    modules = [loadSource(a[0])]
  File "C:\Program Files (x86)\JetBrains\PyCharm Community Edition 143.165\helpers\pycharm\utrunner.py", line 41, in loadSource
    module = imp.load_source(moduleName, fileName)
  File "C:\data\external\pandas\pandas\tests\test_base.py", line 6, in <module>
    import pandas.compat as compat
  File "C:\data\external\pandas\pandas\__init__.py", line 42, in <module>
    import pandas.core.config_init
  File "C:\data\external\pandas\pandas\core\config_init.py", line 17, in <module>
    from pandas.core.format import detect_console_encoding
  File "C:\data\external\pandas\pandas\core\format.py", line 8, in <module>
    from pandas.core.base import PandasObject
  File "C:\data\external\pandas\pandas\core\base.py", line 10, in <module>
    from pandas.core.strings import StringMethods
  File "C:\data\external\pandas\pandas\core\strings.py", line 4, in <module>
    from pandas.core.base import NoNewAttributesMixin
ImportError: cannot import name NoNewAttributesMixin

[jreback]

API change section

[jankatins]

I was actually thinking it should go to bugfix: currently you can't get the value back, so the concept is broken and we only make the error more explicit.

[jreback]

ok (put the .str etal with backticks)

[jankatins]

done (to bug and added backticks)

[jankatins]

BTW: The Series and Indexes are kind of inconsistent:

• Series only has accessors, so no datetime/string/categorical related functions directly exposed under Series but .dt, .cat, which are implemented as Delegate and .str, which is no Delegate
• CategoricalIndex is a Delegate directly (-> CatIndex.codes works directly without CatIndex.cat.codes)
• DatetimeIndex is not a Delegate, but exposes it's methods directly, too, via mixins
• There is no StringIndex but Index.str.... exists (via IndexOpsMixin) for all Indexes?

[jreback]

@JanSchulz

those are not inconsistent, merely implementations. DatetimeIndex, actually defines things (IOW it is the 'target' of the delegation). CategoricalIndex uses the Delegate (so its like Series in this respect). Index/Series are also users of .str.

[jankatins]

I disagree: Consistent would be either:

• Series and Index would use accessors, which are delegates to the real data below (like Series.cat or to a wrapper like in Series.str / Series.dt (ok, preoblematic, because we currently don't have real methods on a string/datetime objects, only numpy/pandas functions. Or the DateTimeIndex, which is used like one...)
• Series and Index would both have subclasses like DateTimeSeries (does not exist) or DateTimeIndex (does exist), which expose the API directly.

[jankatins]

So, move the string part to a new mixin in string.py.. If this is green, I think this should be ready to merge...

[jreback]

@JanSchulz looks good! ping on green.

[jankatins]

Seems travis is dead?!

Ok, I will rebase #11582 on top of this PR, to shortcut that process a bit...

[jreback]

thanks @JanSchulz

go ahead and rebase #11582