BUG: replace should handle Bool blocks correctly

[jreback]

http://stackoverflow.com/questions/21733802/pandas-series-replace-on-boolean-series

by definition they don't have missing values

[TomAugspurger]

Are we ever going to get an integer or bool NaN? I'm haven't followed any developments in this.

[jreback]

its not that difficult if you use a NaT like value (e.g. np.iinfo(np.int64).min)) (which would have to be set for each of the dtypes. (and bool would have to be represented by say np.uint8 or something.

Then of course you would have to go thru lots of routines and deal with the special handling (that is implicity in that you know that int64 if its a dtype of int64 cannot right now have a missing value).

so would have to have someone with lots of time to do this!

up for it @TomAugspurger ?

[TomAugspurger]

I wasn't volunteering :) Maybe someday. I was just asking because I was thinking about skipping int and bool blocks in the interpolation routines.

[jreback]

@TomAugspurger yep....that's makes sense for interpolate for sure

[jreback]

technically replace actually shouldn't ignore on Bool blocks but only replace if the values are True/False

[jreback]

@cpcloud want 2 take this?

[cpcloud]

Happy to fix the bool issue here. Not going to implement inans or bool nans

[jreback]

hah!

that's what I meant :)

FYI also need to test case where u are replacing say True with a value (which prob will switch dtype)

[cpcloud]

weird just those no-op cases fail ... the other cases work

[cpcloud]

into the cave of replace we go 😄

[cpcloud]

@jreback Turns out it's not an issue with blocks or replace per se. In com.mask_missing string values to be masked are coerced to bool. If the replacement strings are non-empty then the mask is the original bool array. That's then passed to putmask eventually which of course replaces those True values with the string