Chang She created ARROW-18273:
---------------------------------
Summary: For extension types, compute kernels should default to
storage types?
Key: ARROW-18273
URL: https://issues.apache.org/jira/browse/ARROW-18273
Project: Apache Arrow
Issue Type: Improvement
Components: C++, Python
Affects Versions: 10.0.0
Reporter: Chang She
Currently, compute kernels don't recognize extensions types so that if you were
to define semantic types to indicate things like "this string column is an
image label", you then cannot do things like equals on it.
For example, take the LabelType from
https://github.com/apache/arrow/blob/c3824db8530075e0f8fd26974c193a310003c17a/python/pyarrow/tests/test_extension_type.py
```
In [1]: import pyarrow as pa
In [2]: import pyarrow.compute as pc
In [3]: class LabelType(pa.PyExtensionType):
...:
...: def __init__(self):
...: pa.PyExtensionType.__init__(self, pa.string())
...:
...: def __reduce__(self):
...: return LabelType, ()
...:
In [4]: tbl = pa.Table.from_arrays([pa.ExtensionArray.from_storage(LabelType(),
pa.array(['cat', 'dog', 'person']))], names=['label'])
In [5]: tbl.filter(pc.field('label') == 'cat')
---------------------------------------------------------------------------
ArrowNotImplementedError Traceback (most recent call last)
Cell In [5], line 1
----> 1 tbl.filter(pc.field('label') == 'cat')
File ~/.venv/lance/lib/python3.10/site-packages/pyarrow/table.pxi:2953, in
pyarrow.lib.Table.filter()
File ~/.venv/lance/lib/python3.10/site-packages/pyarrow/_exec_plan.pyx:391, in
pyarrow._exec_plan._filter_table()
File ~/.venv/lance/lib/python3.10/site-packages/pyarrow/_exec_plan.pyx:128, in
pyarrow._exec_plan.execplan()
File ~/.venv/lance/lib/python3.10/site-packages/pyarrow/error.pxi:144, in
pyarrow.lib.pyarrow_internal_check_status()
File ~/.venv/lance/lib/python3.10/site-packages/pyarrow/error.pxi:121, in
pyarrow.lib.check_status()
ArrowNotImplementedError: Function 'equal' has no kernel matching input types
(extension<arrow.py_extension_type<LabelType>>, string)
```
for query systems that push some of the compute down to Arrow (e.g., DuckDB),
it also means that it's much harder for users to work with datasets with
extension types because you don't know which functions will actually work.
Instead, if we can make the compute kernels default to the storage type, it
would make the extension system a lot easier to work with in Arrow.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
