DEPR:change DataFrame.where(..,raise_on_error) -> errors

[m-charlton]

xref #14878

#14878 deprecated the raise_on_error kwarg in favour of the errors kwarg
for DataFrame.astype().

This issue addresses the same deprecation of the raise_on_error kwarg for
DataFrame.where(). Again this should be replaced with the errors kwarg.

The errors kwarg can have the value raise or ignore with a default
value of raise.

This issue can be assigned to me - @m-charlton

[m-charlton]

I'll start on this tomorrow (Wednesday 4th Jan 2017)

[m-charlton]

It looks as if the value of raise_on_error has no effect on the raising of
exceptions for the where method.

According to the docstring for DataFrame.where

raise_on_error : boolean, default True
     Whether to raise on invalid data types (e.g. trying to where on strings)

Then attempting to compare against a string for a DataFrame of say integers,
will result in an exception, this is the case.

For example in the following, raise_on_error is set to 'False'. As I'm
attempting to compare the string a with integer data then I get a 'TypeError'
exception, but the exception should be suppressed and the DataFrame should be
returned as it is in DataFrame.astype()

>>> import pandas as pd
>>> df = pd.DataFrame(data=[1,2,3])
>>> df.where(df == 'a', raise_on_error=False)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../pandas/core/ops.py", line 1290, in f
    res = self._combine_const(other, func, raise_on_error=False)
  File ".../pandas/core/frame.py", line 3634, in _combine_const
    raise_on_error=raise_on_error)
  File ".../pandas/core/internals.py", line 3162, in eval
    return self.apply('eval', **kwargs)
  File ".../pandas/core/internals.py", line 3056, in apply
    applied = getattr(b, f)(**kwargs)
  File ".../pandas/core/internals.py", line 1181, in eval
    repr(other))
TypeError: Could not compare ['a'] with block values

In File .../pandas/core/ops.py. The 'raise_on_error' kwarg is always set to
'False' [L1290]:

1288             # straight boolean comparisions we want to allow all columns
1289             # (regardless of dtype to pass thru) See #4537 for discussion.
1290             res = self._combine_const(other, func, raise_on_error=False)
1291             return res.fillna(True).astype(bool)

This is an aside, the problem seems to be in pandas/core/internals.py

1157         # get the result
1158         try:
1159             with np.errstate(all='ignore'):
1160                 result = get_result(other)
1161 
1162         # if we have an invalid shape/broadcast error
1163         # GH4576, so raise instead of allowing to pass through
1164         except ValueError as detail:
1165             raise
1166         except Exception as detail:
1167             result = handle_error()
1168 
1169         # technically a broadcast error in numpy can 'work' by returning a
1170         # boolean False
1171         if not isinstance(result, np.ndarray):
1172             if not isinstance(result, np.ndarray):
1173 
1174                 # differentiate between an invalid ndarray-ndarray comparison
1175                 # and an invalid type comparison
1176                 if isinstance(values, np.ndarray) and is_list_like(other):
1177                     raise ValueError('Invalid broadcasting comparison [%s] '
1178                                      'with block values' % repr(other))
1179 
1180                 raise TypeError('Could not compare [%s] with block values' %
1181                                 repr(other))

The 'TypeError' is raised at [L1181], but the 'handle_error()' [L1167] is where
the raise_on_error argument is evaluated.

So in this particular case the 'TypeError' exception will always be raised.

Also the expressions on [L1171] & [L1172] are the same which looks like a
mistake.

[jreback]

@m-charlton PR for this?

[jreback]

PR for this?