BUG: tm.assert_frame_equal fails to raise near int64 implementation bounds when check_exact=True

[jbrockmendel]

passing check_exact=True makes it raise as expected, but I was surprised to find it necessary. This likely hides some bugs.

Motivating example adapted from tests.groupby.test_function.test_cummin.

dtype = np.dtype(np.int64)
min_val = np.iinfo(dtype).min

base_df = DataFrame({"A": [1, 1, 1, 1, 2, 2, 2, 2], "B": [3, 4, 3, 2, 2, 3, 2, 1]})
expected_mins = [3, 3, 3, 2, 2, 2, 2, 1]
expected = DataFrame({"B": expected_mins}).astype(dtype)

df = base_df.astype(dtype)

# Test w/ min value for dtype
df.loc[[2, 6], "B"] = min_val
df.loc[[1, 5], "B"] = min_val + 1
result = df.groupby("A").cummin()

expected.loc[[1, 5], "B"] = min_val + 1
expected.loc[[2, 3, 6, 7], "B"] = min_val

tm.assert_frame_equal(result, expected)  #  <-- passes
tm.assert_frame_equal(result, expected, check_exact=True)  # <-- raises


>>> result
                     B
0                    3
1 -9223372036854775808
2 -9223372036854775808
3 -9223372036854775808
4                    2
5 -9223372036854775808
6 -9223372036854775808
7 -9223372036854775808

>>> expected
                     B
0                    3
1 -9223372036854775807
2 -9223372036854775808
3 -9223372036854775808
4                    2
5 -9223372036854775807
6 -9223372036854775808
7 -9223372036854775808

[mzeitlin11]

This ends up falling down to using math.isclose with tolerances large enough that these compare as equal. I'm guessing the default of comparison with tolerance is just to avoid floating point comparison problems. Seems like ints should just always be given an exact comparison by default?

[jbrockmendel]

Seems like ints should just always be given an exact comparison by default?

thats my thought too

[mzeitlin11]

@jbrockmendel how public are these functions (eg can we just change the default for integer data to check_exact=True or do we have to worry about backwards compat?) Also xref #30347

[jbrockmendel]

in the past when ive changed tm.assert_foo functions @jorisvandenbossche eventually reports that a downstream package has been affectd

[topper-123]

IMO the check_exact param should only apply to floats, i.e. the example in OP should fail. I also think this is a bug, i.e. we can fix just this without deprecation.