Subclassed attributes are not serializable

[kjordahl]

In a subclass of pandas objects, pickling an object doesn't serialize properties of an instance of that subclass, even if the attribute has been added to _metadata. It would be hard to override this behavior entirely in the subclass, because it will require updates to __getstate__ and __setstate__, and probably also an addition or subclass of BlockManager. It would be nice if the _metadata serialization was handled in the base pandas class.

Example:

class SubDataFrame(DataFrame):

    _metadata = ['my_data']

    @property
    def _constructor(self):
        return SubDataFrame

sdf = SubDataFrame()
sdf.my_data = 'foo'
sdf.to_pickle('tmp.pkl')
new_sdf = read_pickle('tmp.pkl')
new_sdf.my_data

raises AttributeError: 'SubDataFrame' object has no attribute 'my_data'

(edited original example for correctness)

[jreback]

I think if this were replaces core/generic.py (and no need to have the Series.__getstate__ either)
then this would work (and be back-compat). The reconstruction accepts a dict of these things

 def __getstate__(self):
        meta = dict([ (k, getattr(self, k, None)) for k in self._metadata ])
        return dict(_data=self._data, **meta)

Want to do a pull-request (already lots of tests on prior version pickle recovery in pandas/io/tests/test_pickle.py so that will test the back-compat)

[kjordahl]

@jreback Thanks, that's simpler than I thought. I had to add a couple more attributes (_metadata and _typ) to make sure it gets deserialized. #10557 submitted.