bug in replace?

[jseabold]

I think this is distinct from #5541 though I agree that an API change here might be helpful. AFAICT, I am using replace as 'expected', but all of the keys aren't being matched. Any idea what's going on?

d = {u'fname' :
    {'out_augmented_AUG_2011.json' : pd.Period(year=2011, month=8, freq='M'), 
    'out_augmented_JAN_2011.json' : pd.Period(year=2011, month=1, freq='M'),
    'out_augmented_MAY_2012.json' : pd.Period(year=2012, month=5, freq='M'),
    'out_augmented_SUBSIDY_WEEK.json' : pd.Period(year=2011, month=4, freq='M'),
    'out_augmented_AUG_2012.json' : pd.Period(year=2012, month=8, freq='M'),
    'out_augmented_MAY_2011.json' : pd.Period(year=2011, month=5, freq='M'),
    'out_augmented_SEP_2013.json' : pd.Period(year=2013, month=9, freq='M'),
    }}

df = pd.DataFrame(['out_augmented_AUG_2012.json',
                'out_augmented_SEP_2013.json',
                'out_augmented_SUBSIDY_WEEK.json',
                'out_augmented_MAY_2012.json',
                'out_augmented_MAY_2011.json',
                'out_augmented_AUG_2011.json',
                'out_augmented_JAN_2011.json'], columns=['fname'])

df.replace(d)

On

[31]: pd.version.version
[31]: '0.13.1-753-g4614ac8'

[cpcloud]

this is a dtype issue ... working on it

[cpcloud]

somewhat subtle issue here:

this is happening because Period raises on period == not_a_period rather than returning NotImplemented (as it should). Python itself then does essentially id(x) == id(y) and this returns False. Numpy return False if any error is thrown instead of reraising

[cpcloud]

so what happens then is that in com.mask_missing we get a scalar bool False instead of the vectorized __eq__ when we do np.array(['the_string']) == np.array([Period(...), 'a', 'b', 'c']) and so we keep getting false for every mask comparison thus the only replaced value is the first

[cpcloud]

the problem is that making it conform the proper python semantics is a breaking API change which @jseabold i know is very fond of. e.g., if you were catching a TypeError before, now you'll get a boolean every time no matter what (unless the other's methods are overridden to do something nefarious like raise)

[cpcloud]

there was only a single test for this behavior and that was that it was raising a TypeError on comparison to non periods

[cpcloud]

@jseabold can you give this a spin when you get a chance? thx

[jseabold]

Link?

[cpcloud]

oh duh i haven't put up the PR yet 😑