[ArrayManager] Enable read_parquet to not create 2D blocks when using…

… ArrayManager (pandas-dev#40303)

pandas/core/generic.py

@@ -286,13 +286,18 @@ def _from_mgr(cls, mgr: Manager):
         object.__setattr__(obj, "_attrs", {})
         return obj
 
-    def _as_manager(self: FrameOrSeries, typ: str) -> FrameOrSeries:
+    def _as_manager(
+        self: FrameOrSeries, typ: str, copy: bool_t = True
+    ) -> FrameOrSeries:
         """
         Private helper function to create a DataFrame with specific manager.
 
         Parameters
         ----------
         typ : {"block", "array"}
+        copy : bool, default True
+            Only controls whether the conversion from Block->ArrayManager
+            copies the 1D arrays (to ensure proper/contiguous memory layout).
 
         Returns
         -------
@@ -301,7 +306,7 @@ def _as_manager(self: FrameOrSeries, typ: str) -> FrameOrSeries:
             to be a copy or not.
         """
         new_mgr: Manager
-        new_mgr = mgr_to_mgr(self._mgr, typ=typ)
+        new_mgr = mgr_to_mgr(self._mgr, typ=typ, copy=copy)
         # fastpath of passing a manager doesn't check the option/manager class
         return self._constructor(new_mgr).__finalize__(self)
 

pandas/core/internals/construction.py

@@ -209,10 +209,11 @@ def fill_masked_arrays(data: MaskedRecords, arr_columns: Index) -> list[np.ndarr
     return new_arrays
 
 
-def mgr_to_mgr(mgr, typ: str):
+def mgr_to_mgr(mgr, typ: str, copy: bool = True):
     """
     Convert to specific type of Manager. Does not copy if the type is already
-    correct. Does not guarantee a copy otherwise.
+    correct. Does not guarantee a copy otherwise. `copy` keyword only controls
+    whether conversion from Block->ArrayManager copies the 1D arrays.
     """
     new_mgr: Manager
 
@@ -231,10 +232,15 @@ def mgr_to_mgr(mgr, typ: str):
             new_mgr = mgr
         else:
             if mgr.ndim == 2:
-                arrays = [mgr.iget_values(i).copy() for i in range(len(mgr.axes[0]))]
+                arrays = [mgr.iget_values(i) for i in range(len(mgr.axes[0]))]
+                if copy:
+                    arrays = [arr.copy() for arr in arrays]
                 new_mgr = ArrayManager(arrays, [mgr.axes[1], mgr.axes[0]])
             else:
-                new_mgr = SingleArrayManager([mgr.internal_values()], [mgr.index])
+                array = mgr.internal_values()
+                if copy:
+                    array = array.copy()
+                new_mgr = SingleArrayManager([array], [mgr.index])
     else:
         raise ValueError(f"'typ' needs to be one of {{'block', 'array'}}, got '{typ}'")
     return new_mgr

pandas/io/parquet.py

@@ -231,6 +231,9 @@ def read(
                     "'use_nullable_dtypes=True' is only supported for pyarrow >= 0.16 "
                     f"({self.api.__version__} is installed"
                 )
+        manager = get_option("mode.data_manager")
+        if manager == "array":
+            to_pandas_kwargs["split_blocks"] = True  # type: ignore[assignment]
 
         path_or_handle, handles, kwargs["filesystem"] = _get_path_or_handle(
             path,
@@ -239,9 +242,12 @@ def read(
             mode="rb",
         )
         try:
-            return self.api.parquet.read_table(
+            result = self.api.parquet.read_table(
                 path_or_handle, columns=columns, **kwargs
             ).to_pandas(**to_pandas_kwargs)
+            if manager == "array":
+                result = result._as_manager("array", copy=False)
+            return result
         finally:
             if handles is not None:
                 handles.close()

pandas/tests/io/test_parquet.py

@@ -9,6 +9,8 @@
 import numpy as np
 import pytest
 
+from pandas._config import get_option
+
 from pandas.compat import (
     PY38,
     is_platform_windows,
@@ -41,20 +43,21 @@
     _HAVE_FASTPARQUET = False
 
 
-pytestmark = [
-    pytest.mark.filterwarnings("ignore:RangeIndex.* is deprecated:DeprecationWarning"),
-    # TODO(ArrayManager) fastparquet / pyarrow rely on BlockManager internals
-    td.skip_array_manager_not_yet_implemented,
-]
+pytestmark = pytest.mark.filterwarnings(
+    "ignore:RangeIndex.* is deprecated:DeprecationWarning"
+)
+
 
+# TODO(ArrayManager) fastparquet relies on BlockManager internals
 
 # setup engines & skips
 @pytest.fixture(
     params=[
         pytest.param(
             "fastparquet",
             marks=pytest.mark.skipif(
-                not _HAVE_FASTPARQUET, reason="fastparquet is not installed"
+                not _HAVE_FASTPARQUET or get_option("mode.data_manager") == "array",
+                reason="fastparquet is not installed or ArrayManager is used",
             ),
         ),
         pytest.param(
@@ -80,6 +83,8 @@ def pa():
 def fp():
     if not _HAVE_FASTPARQUET:
         pytest.skip("fastparquet is not installed")
+    elif get_option("mode.data_manager") == "array":
+        pytest.skip("ArrayManager is not supported with fastparquet")
     return "fastparquet"
 
 
@@ -923,6 +928,18 @@ def test_filter_row_groups(self, pa):
             )
         assert len(result) == 1
 
+    def test_read_parquet_manager(self, pa, using_array_manager):
+        # ensure that read_parquet honors the pandas.options.mode.data_manager option
+        df = pd.DataFrame(np.random.randn(10, 3), columns=["A", "B", "C"])
+
+        with tm.ensure_clean() as path:
+            df.to_parquet(path, pa)
+            result = read_parquet(path, pa)
+        if using_array_manager:
+            assert isinstance(result._mgr, pd.core.internals.ArrayManager)
+        else:
+            assert isinstance(result._mgr, pd.core.internals.BlockManager)
+
 
 class TestParquetFastParquet(Base):
     def test_basic(self, fp, df_full):