Backport PR #39239: DEPR: raise deprecation warning in numpy ufuncs on DataFrames if not aligned + fallback to <1.2.0 behaviour (#39288)

Co-authored-by: Joris Van den Bossche <[email protected]>

Author: meeseeksmachine

doc/source/whatsnew/v1.2.0.rst

@@ -286,6 +286,8 @@ Other enhancements
 - Added methods :meth:`IntegerArray.prod`, :meth:`IntegerArray.min`, and :meth:`IntegerArray.max` (:issue:`33790`)
 - Calling a NumPy ufunc on a ``DataFrame`` with extension types now preserves the extension types when possible (:issue:`23743`)
 - Calling a binary-input NumPy ufunc on multiple ``DataFrame`` objects now aligns, matching the behavior of binary operations and ufuncs on ``Series`` (:issue:`23743`).
+  This change has been reverted in pandas 1.2.1, and the behaviour to not align DataFrames
+  is deprecated instead, see the :ref:`the 1.2.1 release notes <whatsnew_121.ufunc_deprecation>`.
 - Where possible :meth:`RangeIndex.difference` and :meth:`RangeIndex.symmetric_difference` will return :class:`RangeIndex` instead of :class:`Int64Index` (:issue:`36564`)
 - :meth:`DataFrame.to_parquet` now supports :class:`MultiIndex` for columns in parquet format (:issue:`34777`)
 - :func:`read_parquet` gained a ``use_nullable_dtypes=True`` option to use nullable dtypes that use ``pd.NA`` as missing value indicator where possible for the resulting DataFrame (default is ``False``, and only applicable for ``engine="pyarrow"``) (:issue:`31242`)
@@ -536,6 +538,14 @@ Deprecations
 - The ``inplace`` parameter of :meth:`Categorical.remove_unused_categories` is deprecated and will be removed in a future version (:issue:`37643`)
 - The ``null_counts`` parameter of :meth:`DataFrame.info` is deprecated and replaced by ``show_counts``. It will be removed in a future version (:issue:`37999`)
 
+**Calling NumPy ufuncs on non-aligned DataFrames**
+
+Calling NumPy ufuncs on non-aligned DataFrames changed behaviour in pandas
+1.2.0 (to align the inputs before calling the ufunc), but this change is
+reverted in pandas 1.2.1. The behaviour to not align is now deprecated instead,
+see the :ref:`the 1.2.1 release notes <whatsnew_121.ufunc_deprecation>` for
+more details.
+
 .. ---------------------------------------------------------------------------
 
 

doc/source/whatsnew/v1.2.1.rst

@@ -1,6 +1,6 @@
 .. _whatsnew_121:
 
-What's new in 1.2.1 (January 18, 2021)
+What's new in 1.2.1 (January 20, 2021)
 --------------------------------------
 
 These are the changes in pandas 1.2.1. See :ref:`release` for a full changelog
@@ -42,6 +42,79 @@ As a result, bugs reported as fixed in pandas 1.2.0 related to inconsistent tick
 
 .. ---------------------------------------------------------------------------
 
+.. _whatsnew_121.ufunc_deprecation:
+
+Calling NumPy ufuncs on non-aligned DataFrames
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Before pandas 1.2.0, calling a NumPy ufunc on non-aligned DataFrames (or
+DataFrame / Series combination) would ignore the indices, only match
+the inputs by shape, and use the index/columns of the first DataFrame for
+the result:
+
+.. code-block:: python
+
+    >>> df1 = pd.DataFrame({"a": [1, 2], "b": [3, 4]}, index=[0, 1])
+    ... df2 = pd.DataFrame({"a": [1, 2], "b": [3, 4]}, index=[1, 2])
+    >>> df1
+       a  b
+    0  1  3
+    1  2  4
+    >>> df2
+       a  b
+    1  1  3
+    2  2  4
+
+    >>> np.add(df1, df2)
+       a  b
+    0  2  6
+    1  4  8
+
+This contrasts with how other pandas operations work, which first align
+the inputs:
+
+.. code-block:: python
+
+    >>> df1 + df2
+         a    b
+    0  NaN  NaN
+    1  3.0  7.0
+    2  NaN  NaN
+
+In pandas 1.2.0, we refactored how NumPy ufuncs are called on DataFrames, and
+this started to align the inputs first (:issue:`39184`), as happens in other
+pandas operations and as it happens for ufuncs called on Series objects.
+
+For pandas 1.2.1, we restored the previous behaviour to avoid a breaking
+change, but the above example of ``np.add(df1, df2)`` with non-aligned inputs
+will now to raise a warning, and a future pandas 2.0 release will start
+aligning the inputs first (:issue:`39184`). Calling a NumPy ufunc on Series
+objects (eg ``np.add(s1, s2)``) already aligns and continues to do so.
+
+To avoid the warning and keep the current behaviour of ignoring the indices,
+convert one of the arguments to a NumPy array:
+
+.. code-block:: python
+
+    >>> np.add(df1, np.asarray(df2))
+       a  b
+    0  2  6
+    1  4  8
+
+To obtain the future behaviour and silence the warning, you can align manually
+before passing the arguments to the ufunc:
+
+.. code-block:: python
+
+    >>> df1, df2 = df1.align(df2)
+    >>> np.add(df1, df2)
+         a    b
+    0  NaN  NaN
+    1  3.0  7.0
+    2  NaN  NaN
+
+.. ---------------------------------------------------------------------------
+
 .. _whatsnew_121.bug_fixes:
 
 Bug fixes

pandas/core/arraylike.py

@@ -149,6 +149,85 @@ def __rpow__(self, other):
         return self._arith_method(other, roperator.rpow)
 
 
+# -----------------------------------------------------------------------------
+# Helpers to implement __array_ufunc__
+
+
+def _is_aligned(frame, other):
+    """
+    Helper to check if a DataFrame is aligned with another DataFrame or Series.
+    """
+    from pandas import DataFrame
+
+    if isinstance(other, DataFrame):
+        return frame._indexed_same(other)
+    else:
+        # Series -> match index
+        return frame.columns.equals(other.index)
+
+
+def _maybe_fallback(ufunc: Callable, method: str, *inputs: Any, **kwargs: Any):
+    """
+    In the future DataFrame, inputs to ufuncs will be aligned before applying
+    the ufunc, but for now we ignore the index but raise a warning if behaviour
+    would change in the future.
+    This helper detects the case where a warning is needed and then fallbacks
+    to applying the ufunc on arrays to avoid alignment.
+
+    See https://github.com/pandas-dev/pandas/pull/39239
+    """
+    from pandas import DataFrame
+    from pandas.core.generic import NDFrame
+
+    n_alignable = sum(isinstance(x, NDFrame) for x in inputs)
+    n_frames = sum(isinstance(x, DataFrame) for x in inputs)
+
+    if n_alignable >= 2 and n_frames >= 1:
+        # if there are 2 alignable inputs (Series or DataFrame), of which at least 1
+        # is a DataFrame -> we would have had no alignment before -> warn that this
+        # will align in the future
+
+        # the first frame is what determines the output index/columns in pandas < 1.2
+        first_frame = next(x for x in inputs if isinstance(x, DataFrame))
+
+        # check if the objects are aligned or not
+        non_aligned = sum(
+            not _is_aligned(first_frame, x) for x in inputs if isinstance(x, NDFrame)
+        )
+
+        # if at least one is not aligned -> warn and fallback to array behaviour
+        if non_aligned:
+            warnings.warn(
+                "Calling a ufunc on non-aligned DataFrames (or DataFrame/Series "
+                "combination). Currently, the indices are ignored and the result "
+                "takes the index/columns of the first DataFrame. In the future , "
+                "the DataFrames/Series will be aligned before applying the ufunc.\n"
+                "Convert one of the arguments to a NumPy array "
+                "(eg 'ufunc(df1, np.asarray(df2)') to keep the current behaviour, "
+                "or align manually (eg 'df1, df2 = df1.align(df2)') before passing to "
+                "the ufunc to obtain the future behaviour and silence this warning.",
+                FutureWarning,
+                stacklevel=4,
+            )
+
+            # keep the first dataframe of the inputs, other DataFrame/Series is
+            # converted to array for fallback behaviour
+            new_inputs = []
+            for x in inputs:
+                if x is first_frame:
+                    new_inputs.append(x)
+                elif isinstance(x, NDFrame):
+                    new_inputs.append(np.asarray(x))
+                else:
+                    new_inputs.append(x)
+
+            # call the ufunc on those transformed inputs
+            return getattr(ufunc, method)(*new_inputs, **kwargs)
+
+    # signal that we didn't fallback / execute the ufunc yet
+    return NotImplemented
+
+
 def array_ufunc(self, ufunc: Callable, method: str, *inputs: Any, **kwargs: Any):
     """
     Compatibility with numpy ufuncs.
@@ -162,6 +241,11 @@ def array_ufunc(self, ufunc: Callable, method: str, *inputs: Any, **kwargs: Any)
 
     cls = type(self)
 
+    # for backwards compatibility check and potentially fallback for non-aligned frames
+    result = _maybe_fallback(ufunc, method, *inputs, **kwargs)
+    if result is not NotImplemented:
+        return result
+
     # for binary ops, use our custom dunder methods
     result = maybe_dispatch_ufunc_to_dunder_op(self, ufunc, method, *inputs, **kwargs)
     if result is not NotImplemented:

pandas/tests/frame/test_ufunc.py

@@ -1,6 +1,8 @@
 import numpy as np
 import pytest
 
+import pandas.util._test_decorators as td
+
 import pandas as pd
 import pandas._testing as tm
 
@@ -70,12 +72,19 @@ def test_binary_input_aligns_columns(dtype_a, dtype_b):
         dtype_b["C"] = dtype_b.pop("B")
 
     df2 = pd.DataFrame({"A": [1, 2], "C": [3, 4]}).astype(dtype_b)
-    result = np.heaviside(df1, df2)
-    expected = np.heaviside(
-        np.array([[1, 3, np.nan], [2, 4, np.nan]]),
-        np.array([[1, np.nan, 3], [2, np.nan, 4]]),
-    )
-    expected = pd.DataFrame(expected, index=[0, 1], columns=["A", "B", "C"])
+    with tm.assert_produces_warning(FutureWarning):
+        result = np.heaviside(df1, df2)
+    # Expected future behaviour:
+    # expected = np.heaviside(
+    #     np.array([[1, 3, np.nan], [2, 4, np.nan]]),
+    #     np.array([[1, np.nan, 3], [2, np.nan, 4]]),
+    # )
+    # expected = pd.DataFrame(expected, index=[0, 1], columns=["A", "B", "C"])
+    expected = pd.DataFrame([[1.0, 1.0], [1.0, 1.0]], columns=["A", "B"])
+    tm.assert_frame_equal(result, expected)
+
+    # ensure the expected is the same when applying with numpy array
+    result = np.heaviside(df1, df2.values)
     tm.assert_frame_equal(result, expected)
 
 
@@ -85,23 +94,35 @@ def test_binary_input_aligns_index(dtype):
         pytest.xfail(reason="Extension / mixed with multiple inputs not implemented.")
     df1 = pd.DataFrame({"A": [1, 2], "B": [3, 4]}, index=["a", "b"]).astype(dtype)
     df2 = pd.DataFrame({"A": [1, 2], "B": [3, 4]}, index=["a", "c"]).astype(dtype)
-    result = np.heaviside(df1, df2)
-    expected = np.heaviside(
-        np.array([[1, 3], [3, 4], [np.nan, np.nan]]),
-        np.array([[1, 3], [np.nan, np.nan], [3, 4]]),
+    with tm.assert_produces_warning(FutureWarning):
+        result = np.heaviside(df1, df2)
+    # Expected future behaviour:
+    # expected = np.heaviside(
+    #     np.array([[1, 3], [3, 4], [np.nan, np.nan]]),
+    #     np.array([[1, 3], [np.nan, np.nan], [3, 4]]),
+    # )
+    # # TODO(FloatArray): this will be Float64Dtype.
+    # expected = pd.DataFrame(expected, index=["a", "b", "c"], columns=["A", "B"])
+    expected = pd.DataFrame(
+        [[1.0, 1.0], [1.0, 1.0]], columns=["A", "B"], index=["a", "b"]
     )
-    # TODO(FloatArray): this will be Float64Dtype.
-    expected = pd.DataFrame(expected, index=["a", "b", "c"], columns=["A", "B"])
+    tm.assert_frame_equal(result, expected)
+
+    # ensure the expected is the same when applying with numpy array
+    result = np.heaviside(df1, df2.values)
     tm.assert_frame_equal(result, expected)
 
 
+@pytest.mark.filterwarnings("ignore:Calling a ufunc on non-aligned:FutureWarning")
 def test_binary_frame_series_raises():
     # We don't currently implement
     df = pd.DataFrame({"A": [1, 2]})
-    with pytest.raises(NotImplementedError, match="logaddexp"):
+    # with pytest.raises(NotImplementedError, match="logaddexp"):
+    with pytest.raises(ValueError, match=""):
         np.logaddexp(df, df["A"])
 
-    with pytest.raises(NotImplementedError, match="logaddexp"):
+    # with pytest.raises(NotImplementedError, match="logaddexp"):
+    with pytest.raises(ValueError, match=""):
         np.logaddexp(df["A"], df)
 
 
@@ -130,3 +151,92 @@ def test_frame_outer_deprecated():
     df = pd.DataFrame({"A": [1, 2]})
     with tm.assert_produces_warning(FutureWarning):
         np.subtract.outer(df, df)
+
+
+def test_alignment_deprecation():
+    # https://github.com/pandas-dev/pandas/issues/39184
+    df1 = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6]})
+    df2 = pd.DataFrame({"b": [1, 2, 3], "c": [4, 5, 6]})
+    s1 = pd.Series([1, 2], index=["a", "b"])
+    s2 = pd.Series([1, 2], index=["b", "c"])
+
+    # binary dataframe / dataframe
+    expected = pd.DataFrame({"a": [2, 4, 6], "b": [8, 10, 12]})
+
+    with tm.assert_produces_warning(None):
+        # aligned -> no warning!
+        result = np.add(df1, df1)
+    tm.assert_frame_equal(result, expected)
+
+    with tm.assert_produces_warning(FutureWarning):
+        # non-aligned -> warns
+        result = np.add(df1, df2)
+    tm.assert_frame_equal(result, expected)
+
+    result = np.add(df1, df2.values)
+    tm.assert_frame_equal(result, expected)
+
+    result = np.add(df1.values, df2)
+    expected = pd.DataFrame({"b": [2, 4, 6], "c": [8, 10, 12]})
+    tm.assert_frame_equal(result, expected)
+
+    # binary dataframe / series
+    expected = pd.DataFrame({"a": [2, 3, 4], "b": [6, 7, 8]})
+
+    with tm.assert_produces_warning(None):
+        # aligned -> no warning!
+        result = np.add(df1, s1)
+    tm.assert_frame_equal(result, expected)
+
+    with tm.assert_produces_warning(FutureWarning):
+        result = np.add(df1, s2)
+    tm.assert_frame_equal(result, expected)
+
+    with tm.assert_produces_warning(FutureWarning):
+        result = np.add(s2, df1)
+    tm.assert_frame_equal(result, expected)
+
+    result = np.add(df1, s2.values)
+    tm.assert_frame_equal(result, expected)
+
+
+@td.skip_if_no("numba", "0.46.0")
+def test_alignment_deprecation_many_inputs():
+    # https://github.com/pandas-dev/pandas/issues/39184
+    # test that the deprecation also works with > 2 inputs -> using a numba
+    # written ufunc for this because numpy itself doesn't have such ufuncs
+    from numba import float64, vectorize
+
+    @vectorize([float64(float64, float64, float64)])
+    def my_ufunc(x, y, z):
+        return x + y + z
+
+    df1 = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6]})
+    df2 = pd.DataFrame({"b": [1, 2, 3], "c": [4, 5, 6]})
+    df3 = pd.DataFrame({"a": [1, 2, 3], "c": [4, 5, 6]})
+
+    with tm.assert_produces_warning(FutureWarning):
+        result = my_ufunc(df1, df2, df3)
+    expected = pd.DataFrame([[3.0, 12.0], [6.0, 15.0], [9.0, 18.0]], columns=["a", "b"])
+    tm.assert_frame_equal(result, expected)
+
+    # all aligned -> no warning
+    with tm.assert_produces_warning(None):
+        result = my_ufunc(df1, df1, df1)
+    tm.assert_frame_equal(result, expected)
+
+    # mixed frame / arrays
+    with tm.assert_produces_warning(FutureWarning):
+        result = my_ufunc(df1, df2, df3.values)
+    tm.assert_frame_equal(result, expected)
+
+    # single frame -> no warning
+    with tm.assert_produces_warning(None):
+        result = my_ufunc(df1, df2.values, df3.values)
+    tm.assert_frame_equal(result, expected)
+
+    # takes indices of first frame
+    with tm.assert_produces_warning(FutureWarning):
+        result = my_ufunc(df1.values, df2, df3)
+    expected = expected.set_axis(["b", "c"], axis=1)
+    tm.assert_frame_equal(result, expected)