ARROW-1673: [Python] Add support for numpy 'bool' type

[pcmoritz]

This is currently a workaround until the Arrow tensor supports zero copy of byte-length booleans.

[pcmoritz]

I'm not at all sure if this is the right fix; maybe we need a separate field for the width if the type is contained in a tensor? Standardizing around numpy for tensors seems the way to go.

[wesm]

Boolean in Arrow is 1 bit, so don’t make this change. We may need to get creative about dealing with NumPy’s metadata

[pcmoritz]

Fair enough, for our Python serialization we could hack around it by defining a custom serializer. However, for Tensor.from_numpy() we can't do that because the type needs to be fully encoded in the Tensor type. Would it be acceptable to introduce a new type for this? Let me know which solution you prefer.

[pcmoritz]

In particular I'm thinking of introducing a "bool8" type, which is a bool that is encoded as a single byte.

[wesm]

I created https://issues.apache.org/jira/browse/ARROW-1674. This is probably the right way to handle this at the format level. We can separately add a data type in C++. It will be useful to be able to receive numpy.bool_ data with zero copy in Arrow

[robertnishihara]

This just passes bool arrays to the custom serializer, right? Does it make sense to register the custom serializer in the default serialization context or no?

[pcmoritz]

Magically, this is already taken care of. The custom serializer we already have is generic, it will convert the array to nested lists and the custom deserializer will make a numpy array out of it. Not the most efficient solution but it fixes the problem until we have the proper solution :)

[robertnishihara]

Oh, I see. Could be made efficient by having the custom serializer special case bool arrays. But if this is just temporary then no need to.

[pcmoritz]

+1 this is ready to merge as a workaround for ray-project/ray#1121

[wesm]

I haven't looked too deeply, but could you explain how this fix works?

[pcmoritz]

Yeah, the switch case I removed makes it fall back to the default, which uses the custom serializer. This will fall back to the function

arrow/python/pyarrow/serialization.py

Line 81 in a043018

def _serialize_numpy_array(obj):

which converts the array to a nested list upon serialization and back upon deserialization.

[wesm]

Got it, so the workaround is slower / not zero copy. No big deal. I will work to get this fixed more properly + zero copy reads in time for 0.8.0

[wesm]

+1