pandas testsuite with numpy 2.0.0rc1 fails on numexpr

[bnavigator]

I'm currently testing numpy 2 on the the openSUSE python ecosystem. I notice the pandas test suite failing when numpy 2.0.0rc1 is installed instead of 1.26.4:

[  681s] =================================== FAILURES ===================================
[  681s] _ TestTypeCasting.test_binop_typecasting[numexpr-python-left_right0-float64-/] _
[  681s] [gw1] linux -- Python 3.11.8 /usr/bin/python3.11
[  681s] 
[  681s] self = <pandas.tests.computation.test_eval.TestTypeCasting object at 0x7ff16c7666d0>
[  681s] engine = 'numexpr', parser = 'python', op = '/', dt = <class 'numpy.float64'>
[  681s] left_right = ('df', '3')
[  681s] 
[  681s]     @pytest.mark.parametrize("op", ["+", "-", "*", "**", "/"])
[  681s]     # maybe someday... numexpr has too many upcasting rules now
[  681s]     # chain(*(np.core.sctypes[x] for x in ['uint', 'int', 'float']))
[  681s]     @pytest.mark.parametrize("dt", [np.float32, np.float64])
[  681s]     @pytest.mark.parametrize("left_right", [("df", "3"), ("3", "df")])
[  681s]     def test_binop_typecasting(self, engine, parser, op, dt, left_right):
[  681s]         df = DataFrame(np.random.default_rng(2).standard_normal((5, 3)), dtype=dt)
[  681s]         left, right = left_right
[  681s]         s = f"{left} {op} {right}"
[  681s] >       res = pd.eval(s, engine=engine, parser=parser)
[  681s] 
[  681s] pandas/tests/computation/test_eval.py:756: 
[  681s] _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
[  681s] pandas/core/computation/eval.py:357: in eval
[  681s]     ret = eng_inst.evaluate()
[  681s] pandas/core/computation/engines.py:81: in evaluate
[  681s]     res = self._evaluate()
[  681s] pandas/core/computation/engines.py:121: in _evaluate
[  681s]     return ne.evaluate(s, local_dict=scope)
[  681s] /usr/lib64/python3.11/site-packages/numexpr/necompiler.py:977: in evaluate
[  681s]     raise e
[  681s] /usr/lib64/python3.11/site-packages/numexpr/necompiler.py:874: in validate
[  681s]     _names_cache[expr_key] = getExprNames(ex, context, sanitize=sanitize)
[  681s] /usr/lib64/python3.11/site-packages/numexpr/necompiler.py:723: in getExprNames
[  681s]     ex = stringToExpression(text, {}, context, sanitize)
[  681s] _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
[  681s] 
[  681s] s = '(df) / (np.float64(3.0))', types = {}
[  681s] context = {'optimization': 'aggressive', 'truediv': False}, sanitize = True
[  681s] 
[  681s]     def stringToExpression(s, types, context, sanitize: bool=True):
[  681s]         """Given a string, convert it to a tree of ExpressionNode's.
[  681s]         """
[  681s]         # sanitize the string for obvious attack vectors that NumExpr cannot
[  681s]         # parse into its homebrew AST. This is to protect the call to `eval` below.
[  681s]         # We forbid `;`, `:`. `[` and `__`, and attribute access via '.'.
[  681s]         # We cannot ban `.real` or `.imag` however...
[  681s]         # We also cannot ban `.\d*j`, where `\d*` is some digits (or none), e.g. 1.5j, 1.j
[  681s]         if sanitize:
[  681s]             no_whitespace = re.sub(r'\s+', '', s)
[  681s]             skip_quotes = re.sub(r'(\'[^\']*\')', '', no_whitespace)
[  681s]             if _blacklist_re.search(skip_quotes) is not None:
[  681s] >               raise ValueError(f'Expression {s} has forbidden control characters.')
[  681s] E               ValueError: Expression (df) / (np.float64(3.0)) has forbidden control characters.
[  681s] 
[  681s] /usr/lib64/python3.11/site-packages/numexpr/necompiler.py:283: ValueError
...
[  684s] FAILED pandas/tests/computation/test_eval.py::TestTypeCasting::test_binop_typecasting[numexpr-python-left_right0-float64-/]
[  684s] FAILED pandas/tests/computation/test_eval.py::TestTypeCasting::test_binop_typecasting[numexpr-python-left_right1-float64-/]
[  684s] FAILED pandas/tests/computation/test_eval.py::TestTypeCasting::test_binop_typecasting[numexpr-pandas-left_right0-float64-/]
[  684s] FAILED pandas/tests/computation/test_eval.py::TestTypeCasting::test_binop_typecasting[numexpr-pandas-left_right1-float64-/]
[  684s] FAILED pandas/tests/computation/test_eval.py::TestOperations::test_simple_arith_ops[numexpr-python]
[  684s] FAILED pandas/tests/computation/test_eval.py::TestOperations::test_simple_arith_ops[numexpr-pandas]

[FrancescAlted]

This is a consequence of the sanitizers that numexpr implemented a few months ago. In general, it is considered not a good idea to call arbitrary functions inside numexpr expressions, so we encourage to rewrite that test as:

In [11]: ne.evaluate('(df) / b', {'b': np.float64(3.0)})
Out[11]: array([0.33333333, 0.66666667, 1.        ])

[github-actions[bot]]

Message to comment on stale issues. If none provided, will not mark issues stale

[bnavigator]

I am not sure why the pandas bug is still not picked up yet. Numpy as been released for a few weeks already and the regular pandas CI should have encountered the failure by now.

[github-actions[bot]]

Message to comment on stale issues. If none provided, will not mark issues stale

[bnavigator]

This was fixed in pandas main and is being backported in https://github.com/pandas-dev/pandas/pull/59513

[github-actions[bot]]

Message to comment on stale issues. If none provided, will not mark issues stale