BUG: Set check_exact to true if dtype is int

[parthi-siva]

• closes BUG: assert_series_equal not raising on unequal series? #55882
• [Tests added and passed]
• All [code checks passed]
• Added [type annotations]
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

[parthi-siva]

@MarcoGorelli not sure where to add test cases for the change.

Edit: Got it (pandas/tests/util/test_assert_frame_equal.py)

[MarcoGorelli]

Well I'm relieved that CI is still green 😄 Can you update the docs for this method to note that rtol, atol and check_exact don't take effect for int dtype?

Regarding the check - not totally sure about the check, will take a closer look next week

[parthi-siva]

Sure. Updated the documentation. Added note under check_exact. Since we are setting it to True in case of int dtype. Also, documentation for rtol and atol states it is applicable only when check_exact is False so I assumed that it is sufficient to add note under check_exact

[MarcoGorelli]

I think it should always be exact, except for floats?

You should also check for extensions arrays, and include a whatsnew

I've added a commit anyway, hope it's ok - do the changes look alright to you?

[parthi-siva]

Yeah, make sense.

Some testcases are failing in some builds. Is that alright?

[MarcoGorelli]

Looks like another bug was hiding because of the check_exact default 😄 #56136

For now let's xfail the relevant test

[MarcoGorelli]

Looks good to me - although I did add a commit, so leaving open a bit in case others have objections

[parthi-siva]

Sure Marco

[mroeschke]

Thanks @parthi-siva and @MarcoGorelli

[jorisvandenbossche]

This is a different issue than reported in #56131 (there is also no check_dtype=False here)

This doesn't involve different data types (it's all numpy dtypes), and the two objects created here seemingly are the same:

data = np.ma.masked_all((3,), dtype=int)
data[0] = 0
data[1] = 1
data[2] = 2
index = ["a", "b", "c"]
result = Series(data, index=index)
expected = Series([0, 1, 2], index=index, dtype=int)

In [29]: result
Out[29]: 
a    0
b    1
c    2
dtype: int64

In [30]: expected
Out[30]: 
a    0
b    1
c    2
dtype: int64

In [31]: result.dtype == expected.dtype
Out[31]: True

But apparently we have a bug in the Series constructor that preserves the masked array as underlying value if it has no masked elements:

In [32]: result.values
Out[32]: 
masked_array(data=[0, 1, 2],
             mask=[False, False, False],
       fill_value=999999)

In [33]: expected.values
Out[33]: array([0, 1, 2])

That seems like a separate, actual bug we should solve, regardless of the behaviour of check_dtype in assert_series_equal (although being more strict here actually uncovered this bug ..)

[parthi-siva]

Looks like another bug was hiding because of the check_exact default 😄 #56136

For now let's xfail the relevant test

@jorisvandenbossche yeah.. there was another issues because of this