[C++/Dataset] Reading null columns as dictionary not longer possible

[asfimport]

Reading a dataset with a dictionary column where some of the files don't contain any data for that column (and thus are typed as null) broke with #9532. It worked with the 3.0 release though and thus I would consider this a regression.

This can be reproduced using the following Python snippet:

import pyarrow as pa
import pyarrow.parquet as pq
import pyarrow.dataset as ds

table = pa.table({"a": [None, None]})
pq.write_table(table, "test.parquet")
schema = pa.schema([pa.field("a", pa.dictionary(pa.int32(), pa.string()))])
fsds = ds.FileSystemDataset.from_paths(
    paths=["test.parquet"],
    schema=schema,
    format=pa.dataset.ParquetFileFormat(),
    filesystem=pa.fs.LocalFileSystem(),
)
fsds.to_table()

The exception on master is currently:

---------------------------------------------------------------------------
ArrowNotImplementedError                  Traceback (most recent call last)
<ipython-input-14-5f0bc602f16b> in <module>
      6     filesystem=pa.fs.LocalFileSystem(),
      7 )
----> 8 fsds.to_table()

~/Development/arrow/python/pyarrow/_dataset.pyx in pyarrow._dataset.Dataset.to_table()
    456         table : Table instance
    457         """
--> 458         return self._scanner(**kwargs).to_table()
    459 
    460     def head(self, int num_rows, **kwargs):

~/Development/arrow/python/pyarrow/_dataset.pyx in pyarrow._dataset.Scanner.to_table()
   2887             result = self.scanner.ToTable()
   2888 
-> 2889         return pyarrow_wrap_table(GetResultValue(result))
   2890 
   2891     def take(self, object indices):

~/Development/arrow/python/pyarrow/error.pxi in pyarrow.lib.pyarrow_internal_check_status()
    139 cdef api int pyarrow_internal_check_status(const CStatus& status) \
    140         nogil except -1:
--> 141     return check_status(status)
    142 
    143 

~/Development/arrow/python/pyarrow/error.pxi in pyarrow.lib.check_status()
    116             raise ArrowKeyError(message)
    117         elif status.IsNotImplemented():
--> 118             raise ArrowNotImplementedError(message)
    119         elif status.IsTypeError():
    120             raise ArrowTypeError(message)

ArrowNotImplementedError: Unsupported cast from null to dictionary<values=string, indices=int32, ordered=0> (no available cast function for target type)

Reporter: Uwe Korn / @xhochy
Assignee: Krisztian Szucs / @kszucs

PRs and other links:

• GitHub Pull Request #10093

_{Note: This issue was originally created as ARROW-12420. Please see the migration documentation for further details.}

[asfimport]

Uwe Korn / @xhochy:
cc @bkietz  who wrote the PR that broke it ;)

[asfimport]

Krisztian Szucs / @kszucs:
Issue resolved by pull request 10093
#10093

[asfimport]

Uwe Korn / @xhochy:
Thanks @kszucs!