BUG: NaT comparison with numpy array doesn't work element-wise

[jorisvandenbossche]

I noticed the following behaviour:

In [43]: np.array([1, 2, 3]) == pd.Timedelta("1 days")
Out[43]: array([False, False, False])

In [44]: np.array([1, 2, 3]) == pd.Timedelta("nat")
Out[44]: False

@jbrockmendel do you know if this is something we control? (it's similar as the numeric/string comparison with numpy, but here the scalar is one defined by pandas)

[jorisvandenbossche]

Comparing the richcmp implementations of NaT and Timedelta:

pandas/pandas/_libs/tslibs/timedeltas.pyx

Lines 833 to 847 in 36502e9

	elif util.is_array(other):
	# TODO: watch out for zero-dim
	if other.dtype.kind == "m":
	return PyObject_RichCompare(self.asm8, other, op)
	elif other.dtype.kind == "O":
	# operate element-wise
	return np.array(
	[PyObject_RichCompare(self, x, op) for x in other],
	dtype=bool,
	)
	if op == Py_EQ:
	return np.zeros(other.shape, dtype=bool)
	elif op == Py_NE:
	return np.ones(other.shape, dtype=bool)
	return NotImplemented # let other raise TypeError

pandas/pandas/_libs/tslibs/nattype.pyx

Lines 124 to 132 in 36502e9

	elif util.is_array(other):
	if other.dtype.kind in "mM":
	result = np.empty(other.shape, dtype=np.bool_)
	result.fill(_nat_scalar_rules[op])
	elif other.dtype.kind == "O":
	result = np.array([PyObject_RichCompare(self, x, op) for x in other])
	else:
	return NotImplemented
	return result

We indeed explicitly cover this case for Timedelta, but not for NaTType

[jorisvandenbossche]

This appears to be a regression in 1.2.x