Fix DataArray.to_dataframe when the array has MultiIndex (pydata#4442)

Co-authored-by: Keewis <[email protected]>

Author: ghislainp

doc/whats-new.rst

@@ -144,6 +144,9 @@ Bug fixes
   a float64 array (:issue:`4898`, :pull:`4911`). By `Blair Bonnett <https://github.com/bcbnz>`_.
 - Fix decoding of vlen strings using h5py versions greater than 3.0.0 with h5netcdf backend (:issue:`4570`, :pull:`4893`).
   By `Kai Mühlbauer <https://github.com/kmuehlbauer>`_.
+- Allow converting :py:class:`Dataset` or :py:class:`DataArray` objects with a ``MultiIndex``
+  and at least one other dimension to a ``pandas`` object (:issue:`3008`, :pull:`4442`).
+  By `ghislainp <https://github.com/ghislainp>`_.
 
 Documentation
 ~~~~~~~~~~~~~

xarray/core/coordinates.py

@@ -13,6 +13,7 @@
     cast,
 )
 
+import numpy as np
 import pandas as pd
 
 from . import formatting, indexing
@@ -107,8 +108,49 @@ def to_index(self, ordered_dims: Sequence[Hashable] = None) -> pd.Index:
             return self._data.get_index(dim)  # type: ignore
         else:
             indexes = [self._data.get_index(k) for k in ordered_dims]  # type: ignore
-            names = list(ordered_dims)
-            return pd.MultiIndex.from_product(indexes, names=names)
+
+            # compute the sizes of the repeat and tile for the cartesian product
+            # (taken from pandas.core.reshape.util)
+            index_lengths = np.fromiter(
+                (len(index) for index in indexes), dtype=np.intp
+            )
+            cumprod_lengths = np.cumproduct(index_lengths)
+
+            if cumprod_lengths[-1] != 0:
+                # sizes of the repeats
+                repeat_counts = cumprod_lengths[-1] / cumprod_lengths
+            else:
+                # if any factor is empty, the cartesian product is empty
+                repeat_counts = np.zeros_like(cumprod_lengths)
+
+            # sizes of the tiles
+            tile_counts = np.roll(cumprod_lengths, 1)
+            tile_counts[0] = 1
+
+            # loop over the indexes
+            # for each MultiIndex or Index compute the cartesian product of the codes
+
+            code_list = []
+            level_list = []
+            names = []
+
+            for i, index in enumerate(indexes):
+                if isinstance(index, pd.MultiIndex):
+                    codes, levels = index.codes, index.levels
+                else:
+                    code, level = pd.factorize(index)
+                    codes = [code]
+                    levels = [level]
+
+                # compute the cartesian product
+                code_list += [
+                    np.tile(np.repeat(code, repeat_counts[i]), tile_counts[i])
+                    for code in codes
+                ]
+                level_list += levels
+                names += index.names
+
+            return pd.MultiIndex(level_list, code_list, names=names)
 
     def update(self, other: Mapping[Hashable, Any]) -> None:
         other_vars = getattr(other, "variables", other)

xarray/tests/test_dataarray.py

@@ -3635,6 +3635,33 @@ def test_to_dataframe(self):
         with raises_regex(ValueError, "unnamed"):
             arr.to_dataframe()
 
+    def test_to_dataframe_multiindex(self):
+        # regression test for #3008
+        arr_np = np.random.randn(4, 3)
+
+        mindex = pd.MultiIndex.from_product([[1, 2], list("ab")], names=["A", "B"])
+
+        arr = DataArray(arr_np, [("MI", mindex), ("C", [5, 6, 7])], name="foo")
+
+        actual = arr.to_dataframe()
+        assert_array_equal(actual["foo"].values, arr_np.flatten())
+        assert_array_equal(actual.index.names, list("ABC"))
+        assert_array_equal(actual.index.levels[0], [1, 2])
+        assert_array_equal(actual.index.levels[1], ["a", "b"])
+        assert_array_equal(actual.index.levels[2], [5, 6, 7])
+
+    def test_to_dataframe_0length(self):
+        # regression test for #3008
+        arr_np = np.random.randn(4, 0)
+
+        mindex = pd.MultiIndex.from_product([[1, 2], list("ab")], names=["A", "B"])
+
+        arr = DataArray(arr_np, [("MI", mindex), ("C", [])], name="foo")
+
+        actual = arr.to_dataframe()
+        assert len(actual) == 0
+        assert_array_equal(actual.index.names, list("ABC"))
+
     def test_to_pandas_name_matches_coordinate(self):
         # coordinate with same name as array
         arr = DataArray([1, 2, 3], dims="x", name="x")