Repository navigation
[C++][Python] Allow creating a UnionDataset from filtered datasets #55447
Description
Describe the enhancement requested
Trying to create a dataset as a union of filtered datasets currently raises an error:
from pathlib import Path
import tempfile
import pyarrow as pa
import pyarrow.compute as pc
import pyarrow.dataset as ds
with tempfile.TemporaryDirectory() as temp_dir:
temp_path = Path(temp_dir)
t1 = pa.table({"x": [1, 1, 2, 2], "y": [1, 2, 3, 4]})
ds.write_dataset(t1, temp_path / "t1", format="parquet")
ds1 = ds.dataset(temp_path / "t1", format="parquet")
t2 = pa.table({"x": [1, 1, 2, 2], "y": [5, 6, 7, 8]})
ds.write_dataset(t2, temp_path / "t2", format="parquet")
ds2 = ds.dataset(temp_path / "t2", format="parquet")
ds1f = ds1.filter(pc.field("x") > 1)
ds2f = ds2.filter(pc.field("x") < 2)
union = ds.dataset([ds1f, ds2f])Raises:
ValueError: Creating an UnionDataset from filtered or projected Datasets is currently not supported. Union the unfiltered datasets and apply the filter to the resulting union.
The suggested workaround is only valid if the filter applied to each dataset is the same. If the filters are different, or not all datasets are filtered, then there's currently no way to create a dataset that represents the union of these datasets.
This limitation seems to be present because the filter is only stored on the Python Dataset object, but creating a union dataset happens in C++.
Component(s)
C++, Python
Activity
[-]Allow creating a UnionDataset from filtered datasets[/-][+][C++][Python] Allow creating a UnionDataset from filtered datasets[/+]