python create from empty arrow tbl (#2557)

py-polars/polars/convert.py

@@ -144,6 +144,7 @@ def from_dicts(dicts: Sequence[Dict[str, Any]]) -> DataFrame:
     return DataFrame._from_dicts(dicts)
 
 
+# Note that we cannot overload because pyarrow has no stubs :(
 def from_arrow(
     a: Union["pa.Table", "pa.Array", "pa.ChunkedArray"], rechunk: bool = True
 ) -> Union[DataFrame, Series]:

py-polars/polars/internals/construction.py

@@ -437,8 +437,13 @@ def arrow_to_pydf(
             data_dict[name] = column
 
     if len(data_dict) > 0:
-        batches = pa.table(data_dict).to_batches()
-        pydf = PyDataFrame.from_arrow_record_batches(batches)
+        tbl = pa.table(data_dict)
+
+        # path for table without rows that keeps datatype
+        if tbl.shape[0] == 0:
+            pydf = pli.DataFrame._from_pandas(tbl.to_pandas())._df
+        else:
+            pydf = PyDataFrame.from_arrow_record_batches(tbl.to_batches())
     else:
         pydf = pli.DataFrame([])._df
     if rechunk:

py-polars/polars/internals/frame.py

@@ -356,6 +356,7 @@ def _from_pandas(
         -------
         DataFrame
         """
+        # path for table without rows that keeps datatype
         if data.shape[0] == 0:
             return DataFrame([pli.Series(name, data[name]) for name in data.columns])
 

py-polars/tests/test_interop.py

@@ -192,3 +192,9 @@ def test_from_empty_pandas() -> None:
     polars_df = pl.from_pandas(pandas_df)
     assert polars_df.columns == ["A", "fruits"]
     assert polars_df.dtypes == [pl.Float64, pl.Float64]
+
+
+def test_from_empty_arrow() -> None:
+    df = pl.from_arrow(pa.table(pd.DataFrame({"a": [], "b": []})))
+    assert df.columns == ["a", "b"]  # type: ignore
+    assert df.dtypes == [pl.Float64, pl.Float64]  # type: ignore