CI: New Numpydev version broke ci

[phofl]

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the master branch of pandas.

Reproducible Example

Two tests are affected

Issue Description

    @pytest.mark.parametrize("quantile", [0.0, 0.1, 0.45, 0.5, 1])
    @pytest.mark.parametrize(
        "interpolation", ["linear", "lower", "higher", "nearest", "midpoint"]
    )
    @pytest.mark.parametrize(
        "data",
        [
            [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0],
            [8.0, 1.0, 3.0, 4.0, 5.0, 2.0, 6.0, 7.0],
            [0.0, np.nan, 0.2, np.nan, 0.4],
            [np.nan, np.nan, np.nan, np.nan],
            [np.nan, 0.1, np.nan, 0.3, 0.4, 0.5],
            [0.5],
            [np.nan, 0.7, 0.6],
        ],
    )
    def test_rolling_quantile_interpolation_options(quantile, interpolation, data):
        # Tests that rolling window's quantile behavior is analogous to
        # Series' quantile for each interpolation option
        s = Series(data)
    
        q1 = s.quantile(quantile, interpolation)
        q2 = s.expanding(min_periods=1).quantile(quantile, interpolation).iloc[-1]
    
        if np.isnan(q1):
            assert np.isnan(q2)
        else:
>           assert q1 == q2
E           assert 0.040000000000000036 == 0.04000000000000001

pandas/tests/window/test_rolling.py:1664: AssertionError
________ test_rolling_quantile_interpolation_options[data2-linear-0.45] ________
[gw0] linux -- Python 3.9.7 /usr/share/miniconda/envs/pandas-dev/bin/python

quantile = 0.45, interpolation = 'linear', data = [0.0, nan, 0.2, nan, 0.4]

    @pytest.mark.parametrize("quantile", [0.0, 0.1, 0.45, 0.5, 1])
    @pytest.mark.parametrize(
        "interpolation", ["linear", "lower", "higher", "nearest", "midpoint"]
    )
    @pytest.mark.parametrize(
        "data",
        [
            [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0],
            [8.0, 1.0, 3.0, 4.0, 5.0, 2.0, 6.0, 7.0],
            [0.0, np.nan, 0.2, np.nan, 0.4],
            [np.nan, np.nan, np.nan, np.nan],
            [np.nan, 0.1, np.nan, 0.3, 0.4, 0.5],
            [0.5],
            [np.nan, 0.7, 0.6],
        ],
    )
    def test_rolling_quantile_interpolation_options(quantile, interpolation, data):
        # Tests that rolling window's quantile behavior is analogous to
        # Series' quantile for each interpolation option
        s = Series(data)
    
        q1 = s.quantile(quantile, interpolation)
        q2 = s.expanding(min_periods=1).quantile(quantile, interpolation).iloc[-1]
    
        if np.isnan(q1):
            assert np.isnan(q2)
        else:
>           assert q1 == q2
E           assert 0.18000000000000005 == 0.18000000000000002

Expected Behavior

Passing

Installed Versions

INSTALLED VERSIONS

commit : 0861da6
python : 3.8.12.final.0
python-bits : 64
OS : Linux
OS-release : 5.11.0-38-generic
Version : #42~20.04.1-Ubuntu SMP Tue Sep 28 20:41:07 UTC 2021
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_US.UTF-8
LOCALE : en_US.UTF-8

pandas : 1.4.0.dev0+1056.g0861da680b
numpy : 1.22.0.dev0+1693.gf52acb53a
pytz : 2021.1
dateutil : 2.8.2
pip : 21.2.4
setuptools : 58.0.4
Cython : 0.29.24
pytest : 6.2.5
hypothesis : 6.23.1
sphinx : 4.2.0
blosc : None
feather : None
xlsxwriter : 3.0.1
lxml.etree : 4.6.3
html5lib : 1.1
pymysql : None
psycopg2 : None
jinja2 : 3.0.1
IPython : 7.28.0
pandas_datareader: None
bs4 : 4.10.0
bottleneck : 1.3.2
fsspec : 2021.11.0
fastparquet : 0.7.1
gcsfs : 2021.05.0
matplotlib : 3.4.3
numexpr : 2.7.3
odfpy : None
openpyxl : 3.0.9
pandas_gbq : None
pyarrow : 5.0.0
pyxlsb : None
s3fs : 2021.11.0
scipy : 1.7.1
sqlalchemy : 1.4.25
tables : 3.6.1
tabulate : 0.8.9
xarray : 0.18.2
xlrd : 2.0.1
xlwt : 1.3.0
numba : 0.53.1
None

Process finished with exit code 0

[jreback]

cc @seberg

[seberg]

Ping @bzah in case you have an idea. Percentile/quantile was fixed with respect to the interpolation (any input welcome, as mentioned on the mailing list I would like to rename interpolation to `method).

Not quite sure how much we should worry about it, but my guess is that this is due to rounding differences in:

    return n * quantiles + (
             alpha + quantiles * (1 - alpha - beta)
     ) - 1

behaving different for different n in the case where alpha=1 and beta=1 (the case for the default "linear" method).

I was wondering before if we should rewrite it, maybe even using _lerp, but not sure. E.g.: quantiles * (n + 1 - alpha - beta) + (alpha - 1) (not thought through.)

[bzah]

Yes it really seems to come from the rounding of _compute_virtual_index which uses the formula @seberg has given.

You can see that with:

data = [0.0, np.nan, 0.2, np.nan, 0.4]

# the new formula with values replaced 
len(data) * 0.45 + (
            1 + 0.45 * (1 - 1 - 1)
    ) - 1
# Out[20]: 1.7999999999999998

0.45 * (len(data) - 1)
# Out[19]: 1.8

To avoid this issue we could use (n -1) * quantile when the method 7 (a.k.a linear) is used and the full formula for others methods. But it would make the code less readable.

[alimcmaster1]

Unless a fix for this is coming soon on the numpy side? shall we add this build against numpy dev to our allowed failures - since there is quite a CI backlog.

numpy/numpy#19857

[jreback]

yeah let's either xfail that test or allow failures for now

[seberg]

Mid-term, the right thing seems to be either relax the test to allow a few ULPs of error, or modify NumPy. But I guess you are talking about short-term CI-unbreaking?

[jreback]

sure we should fix the test too
but this is making our ci red so let's fix that first

[seberg]

I merged the fixup that ensures we use the simpler (and thus more precise) formula again for the default. That should fix this issue.

A heads up though: I am hoping to rename interpolation= to method= with a DeprecationWarning. This would be very soon and should be in 1.22. If that is problematic, let us know and we could reconsider or take it slower.