Fix array_equal behaviour for masked arrays (#4457)

* fix array_equal behaviour for masked arrays

* address review comment

* fix tests

* fix bug, add test and whatsnew

* address review comments

* Update lib/iris/tests/unit/util/test_array_equal.py

Co-authored-by: Patrick Peglar <[email protected]>

---------

Co-authored-by: Patrick Peglar <[email protected]>

docs/src/whatsnew/latest.rst

@@ -52,6 +52,11 @@ This document explains the changes made to Iris for this release
  :func:`~iris.fileformats.pp.save_pairs_from_cube` because it had no effect.
  (:pull:`5783`)
 
+#. `@stephenworsley`_ made masked arrays on Iris objects now compare as equal
+ precisely when all unmasked points are equal and when the masks are identical.
+ This is due to changes in :func:`~iris.util.array_equal` which previously
+ ignored masks entirely. (:pull:`4457`)
+
 
 🚀 Performance Enhancements
 ===========================

lib/iris/tests/unit/util/test_array_equal.py

@@ -44,11 +44,28 @@ def test_nd(self):
  self.assertTrue(array_equal(array_a, array_b))
  self.assertFalse(array_equal(array_a, array_c))
 
- def test_masked_is_ignored(self):
+ def test_masked_is_not_ignored(self):
  array_a = ma.masked_array([1, 2, 3], mask=[1, 0, 1])
  array_b = ma.masked_array([2, 2, 2], mask=[1, 0, 1])
+ self.assertTrue(array_equal(array_a, array_b))
+
+ def test_masked_is_different(self):
+ array_a = ma.masked_array([1, 2, 3], mask=[1, 0, 1])
+ array_b = ma.masked_array([1, 2, 3], mask=[0, 0, 1])
  self.assertFalse(array_equal(array_a, array_b))
 
+ def test_masked_isnt_unmasked(self):
+ array_a = np.array([1, 2, 2])
+ array_b = ma.masked_array([1, 2, 2], mask=[0, 0, 1])
+ self.assertFalse(array_equal(array_a, array_b))
+
+ def test_masked_unmasked_equivelance(self):
+ array_a = np.array([1, 2, 2])
+ array_b = ma.masked_array([1, 2, 2])
+ array_c = ma.masked_array([1, 2, 2], mask=[0, 0, 0])
+ self.assertTrue(array_equal(array_a, array_b))
+ self.assertTrue(array_equal(array_a, array_c))
+
  def test_fully_masked_arrays(self):
  array_a = ma.masked_array(np.arange(24).reshape(2, 3, 4), mask=True)
  array_b = ma.masked_array(np.arange(24).reshape(2, 3, 4), mask=True)

lib/iris/util.py

@@ -395,8 +395,9 @@ def array_equal(array1, array2, withnans=False):
 
  Notes
  -----
- This provides much the same functionality as :func:`numpy.array_equal`, but
- with additional support for arrays of strings and NaN-tolerant operation.
+ This provides similar functionality to :func:`numpy.array_equal`, but
+ will compare arrays as unequal when their masks differ and has
+ additional support for arrays of strings and NaN-tolerant operation.
 
  This function maintains laziness when called; it does not realise data.
  See more at :doc:`/userguide/real_and_lazy_data`.
@@ -406,17 +407,24 @@ def array_equal(array1, array2, withnans=False):
 
  def normalise_array(array):
  if not is_lazy_data(array):
- array = np.asarray(array)
+ if not ma.isMaskedArray(array):
+ array = np.asanyarray(array)
  return array
 
  array1, array2 = normalise_array(array1), normalise_array(array2)
 
  eq = array1.shape == array2.shape
+ if eq:
+ array1_masked = ma.is_masked(array1)
+ eq = array1_masked == ma.is_masked(array2)
+ if eq and array1_masked:
+ eq = np.array_equal(ma.getmaskarray(array1), ma.getmaskarray(array2))
  if eq:
  eqs = array1 == array2
  if withnans and (array1.dtype.kind == "f" or array2.dtype.kind == "f"):
  eqs = np.where(np.isnan(array1) & np.isnan(array2), True, eqs)
- eq = bool(np.all(eqs))
+ eq = np.all(eqs)
+ eq = bool(eq) or eq is ma.masked
 
  return eq