BUG: combine_first behaves strangely with period index

[kiroh]

If i do:

from pandas import Series, PeriodIndex, Period

pi = PeriodIndex(start=Period('1950-1'), freq='M', end=Period('1950-7'))
a = Series([1,nan,nan,4,5,nan,7], pi)
b = Series([9,9,9,9,9,9,9], pi)

display(a, b, a.combine_first(b))

Then the output is:

1950-01     1
1950-02   NaN
1950-03   NaN
1950-04     4
1950-05     5
1950-06   NaN
1950-07     7
Freq: M

1950-01    9
1950-02    9
1950-03    9
1950-04    9
1950-05    9
1950-06    9
1950-07    9
Freq: M

1930-01   NaN
1930-03   NaN
1930-05   NaN
1930-07   NaN
1930-09   NaN
1930-11   NaN
1931-01   NaN
Freq: M

As you can see the index is changed and the values are all gone.
I think this is a bug. My Pandas version is 0.10.1.

[ghost]

I can't reproduce this with git master, can you try 0.11.0rc1 and see if the problem persists?

[kiroh]

Ok, is there any way to try 0.11.0rc1 without breaking my 0.10.1 installation?

[ghost]

You can use a virtualenv.

[jreback]

@y-p occurs in 0.11.0rc1 too, same result....

[ghost]

Oh my bad, I'm working with an old version. apperently it worked fine in '0.8.0.dev-092fedc'.

[jreback]

looking at code, its not going to work....mark for bug in 0.12 (only affects PeriodIndex), DatetimeIndex works fine

[ghost]

bisected? what broke it?

[jreback]

no idea when broken (obviously >0.8.0<0.10.1)

[cganterh]

It's fixed?

[jreback]

@cganterh sorry...i think got accidently closed

[sinhrks]

Current master raises TypeError: unsupported operand type(s) for +: 'PeriodIndex' and 'PeriodIndex'. #7741 will fix this.