FIX-#7569: Fix handling of pyarrow dtype and empty dataframes (#7570)

Ensure empty Dataframes with pyarrow dtypes maintain the dtype metadata.

Signed-off-by: Michael Peleshenko <[email protected]>

Author: mpeleshenko

modin/core/dataframe/pandas/dataframe/dataframe.py

@@ -174,7 +174,9 @@ def __init__(
         self._row_lengths_cache = row_lengths
         self._column_widths_cache = column_widths
         self._pandas_backend = pandas_backend
-        if pandas_backend != "pyarrow":
+        if pandas_backend != "pyarrow" or len(partitions) == 0:
+            # If the backend is pyarrow and there are no partitions, the computed dtype otherwise becomes NaN,
+            # which means we lost the dtype, so actually set it in that case
             self.set_dtypes_cache(dtypes)
         else:
             # In this case, the type precomputation may be incorrect; we need

modin/tests/pandas/dataframe/test_default.py

@@ -18,6 +18,7 @@
 import numpy as np
 import pandas
 import pandas._libs.lib as lib
+import pyarrow as pa
 import pytest
 from numpy.testing import assert_array_equal
 
@@ -1533,3 +1534,9 @@ def test_series_does_not_warn_distributing_takes_time():
     with warnings.catch_warnings():
         warnings.filterwarnings("error", regex, UserWarning)
         pd.Series(np.random.randint(1_000_000, size=(2_400_000)))
+
+
+@pytest.mark.parametrize("dtype", [np.int64, pd.ArrowDtype(pa.int64())])
+def test_empty_df_dtypes(dtype):
+    df = pd.DataFrame({"A": []}, dtype=dtype)
+    assert df.dtypes["A"] == dtype