Consider adding `__hash__` to `GeometryExtensionType`

[FBruzzesi]

TL;DR from an issue spotted in narwhals (narwhals-dev/narwhals#3222).

Would you consider making the extension type hashable? We rely quite heavily on caching (via lru_cache) for mapping back and forth between datatypes.

This logic breaks if a pyarrow table has an (extension) dtype that does not implement an __hash__ method.

[paleolimbot]

Sure!

[dangotbanned]

@paleolimbot I feel like I'm missing something here re (narwhals-dev/narwhals#3222 (comment))

I would have expected the pyarrow.DataType.__hash__ implementation to have handled this for you:

https://github.com/apache/arrow/blob/bab558061696ddc1841148d6210424b12923d48e/python/pyarrow/types.pxi#L355-L356

    def __hash__(self):
        return hash(str(self))

Or if there were some other requirement that pyarrow has to determine hashability, then the error here wouldn't be coming from the interpreter:

import pyarrow as pa

hash(pa.UnknownExtensionType(pa.int64(), b"serialized"))
TypeError: unhashable type: 'pyarrow.lib.UnknownExtensionType'

Updated

Okay found the issue

Mentioned in the python data model docs object.__hash__:

If a class that overrides __eq__() needs to retain the implementation of __hash__() from a parent class, the interpreter must be told this explicitly by setting __hash__ = <ParentClass>.__hash__.

That does happen in ExtensionType:

https://github.com/apache/arrow/blob/bab558061696ddc1841148d6210424b12923d48e/python/pyarrow/types.pxi#L1874-L1882