hypothesis.extra.pandas.column, errors out when elements = lists and unique = True

[chrisli-livongo]

When defining a pandas.columns with lists

from hypothesis.extra.pandas import column
from hypothesis.strategies import lists

column(name='group_array', elements=lists(integers(min_value=-2147483648, max_value=2147483647), max_size=5, min_size=5, unique=True), dtype=list, fill=None, unique=True)

We got this error

...
/usr/local/lib/python3.8/dist-packages/hypothesis/internal/conjecture/data.py:884: in draw
    return strategy.do_draw(self)
/usr/local/lib/python3.8/dist-packages/hypothesis/strategies/_internal/lazy.py:168: in do_draw
    return data.draw(self.wrapped_strategy)
/usr/local/lib/python3.8/dist-packages/hypothesis/internal/conjecture/data.py:884: in draw
    return strategy.do_draw(self)
/usr/local/lib/python3.8/dist-packages/hypothesis/strategies/_internal/lazy.py:168: in do_draw
    return data.draw(self.wrapped_strategy)
/usr/local/lib/python3.8/dist-packages/hypothesis/internal/conjecture/data.py:884: in draw
    return strategy.do_draw(self)
/usr/local/lib/python3.8/dist-packages/hypothesis/strategies/_internal/core.py:1393: in do_draw
    return self.definition(data.draw, *self.args, **self.kwargs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

draw = <bound method ConjectureData.draw of ConjectureData(VALID, 1961 bytes, frozen)>

    @st.composite
    def just_draw_columns(draw):
        index = draw(index_strategy)
        local_index_strategy = st.just(index)
    
        data = OrderedDict((c.name, None) for c in rewritten_columns)
    
        # Depending on how the columns are going to be generated we group
        # them differently to get better shrinking. For columns with fill
        # enabled, the elements can be shrunk independently of the size,
        # so we can just shrink by shrinking the index then shrinking the
        # length and are generally much more free to move data around.
    
        # For columns with no filling the problem is harder, and drawing
        # them like that would result in rows being very far apart from
        # each other in the underlying data stream, which gets in the way
        # of shrinking. So what we do is reorder and draw those columns
        # row wise, so that the values of each row are next to each other.
        # This makes life easier for the shrinker when deleting blocks of
        # data.
        columns_without_fill = [c for c in rewritten_columns if c.fill.is_empty]
    
        if columns_without_fill:
            for c in columns_without_fill:
                data[c.name] = pandas.Series(
                    np.zeros(shape=len(index), dtype=c.dtype), index=index
                )
            seen = {c.name: set() for c in columns_without_fill if c.unique}
    
            for i in range(len(index)):
                for c in columns_without_fill:
                    if c.unique:
                        for _ in range(5):
                            value = draw(c.elements)
>                           if value not in seen[c.name]:
E                           TypeError: unhashable type: 'numpy.ndarray'

/usr/local/lib/python3.8/dist-packages/hypothesis/extra/pandas/impl.py:578: TypeError

But it works fine, when we set column unique=False:
column(name='group_array', elements=lists(integers(min_value=-2147483648, max_value=2147483647), max_size=5, min_size=5, unique=True), dtype=list, fill=None, unique=False)

[honno]

Have no idea myself sorry, here's just a reproducer for folks.

>>> from hypothesis import strategies as st
>>> from hypothesis.extra import pandas as pdst
>>> col1 = pdst.column(elements=st.lists(st.integers()), dtype=list, unique=False)
>>> pdst.data_frames([col1]).example()
    0
0  []
>>> col2 = pdst.column(elements=st.lists(st.integers()), dtype=list, unique=True)
>>> pdst.data_frames([col2]).example()
TypeError: unhashable type: 'numpy.ndarray'

[Zac-HD]

Well, the problem is just that we track a dict[str, set] of seen items for each column name, but if the generated element is not hashable (and indeed ndarray is not) then we get a type error.

You get the same error from st.lists(st.lists(st.none()), unique=True).example(), so I think this is a docs issue rather than a bug.

[rsokl]

I had labeled this a bug because I think that pdst.column(elements=st.lists(st.integers()), dtype=list, unique=True) should immediately raise since it is not supported.

Even if the user has insight into the fact that we require "hashability" in order to draw unique elements, and used pdst.column(elements=st.lists(st.integers()).map(tuple), dtype=tuple, unique=True) instead, they would still hit the same error because pdst.data_frames will convert the sequence to an ndarray under the hood, which of course is not hashable.

[Zac-HD]

The underlying problem here is that there's no such thing as dtype=list or dtype=tuple; both are aliases of dtype=object. So the only thing we can really do is emit a warning when we see this and suspect a misunderstanding!