opt out of bottleneck for nanmean

[sebasv]

• closes BUG: np.mean(pd.Series) != np.mean(pd.Series.values) #42878
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

[pep8speaks]

Hello @sebasv! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2022-07-17 13:11:07 UTC

[sebasv]

I added 2 unit tests to validate that numerical precision is bounded by log(n). If anyone has suggestions how to improve these tests let me know

[mroeschke]

Confirming that this should be disabled for integer types too? i.e. the precision loss is similar for integers and therefore should also be disabled?

[sebasv]

@mroeschke I will verify this. Addition is pure for integers so depending on the bottleneck implementation it should be almost exact

[JMBurley]

Glad to see this update. Should it be targeting pandas 1.5.0 or 1.4.4 release?

[mroeschke]

Glad to see this update. Should it be targeting pandas 1.5.0 or 1.4.4 release?

1.5. Point releases are reserved for regression in behavior from recent releases.

[sebasv]

Integer mean should also be opted out of. In the case of integer arrays it seems that bottleneck accumulates the sum in a float64 type, wich means that the problem persists there as well. I'm adding int types to the unit test as well

Example

>>> a = np.full((1_000_000_000, 1), 1_000_000_000_000, dtype=np.int64)
>>> bottleneck.nanmean(a)
1000000022905.4894
>>> np.nanmean(a)
1000000000000.0

[mroeschke]

Thanks @sebasv and @JMBurley for review