Issue warning in slow ExtensionArray base methods

[TomAugspurger]

In the ExtensionArray docstring, we note that

    Some methods require casting the ExtensionArray to an ndarray of Python
    objects with ``self.astype(object)``, which may be expensive. When
    performance is a concern, we highly recommend overriding the following
    methods:

    * fillna
    * dropna
    * unique
    * factorize / _values_for_factorize
    * argsort / _values_for_argsort

I think we should also issue a warning in those cases. We would filter the warning by default, so it doesn't appear to users, but would show up when the base tests are being run. Then the EA author can choose to ignore that warning, or implement it more efficiently.

[jreback]

sounds reasonable to issue a PerformanceWarning but don’t want it to show for end users

[TomAugspurger]

It'll be a new warning class though. We can't reuse `pandas.PerformanceWarning` since we'll add a filter for the ExtensionArray one to hide it by default.

…

On Wed, Dec 26, 2018 at 6:50 AM Jeff Reback ***@***.***> wrote: sounds reasonable to issue a PerformanceWarning but don’t want it to show for end users — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#24433 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIiiuhfbjP1rxpLbnjtM9A4KORKXKks5u83CGgaJpZM4ZhtvC> .

[TomAugspurger]

I may not get to this for 0.24.

https://github.com/pandas-dev/pandas/compare/master...TomAugspurger:astype-warning?expand=1 is basically done, but it exposed some troubles with our current approach to parametrizing tests. Basically, if a parametrized test was defined directly on a fixture, I had to duplicate the parametrization in the subclass, which is no good.

That branch has started moving those parametrization to class attributes, e.g.
https://github.com/pandas-dev/pandas/compare/master...TomAugspurger:astype-warning?expand=1#diff-9bd888be54b675f517a4ee682a58cbe9R88. If anyone has a cleaner alternative I'd be happy to hear it.

[jbrockmendel]

Tried this out and got 500ish test failures for uncaught warnings. A little more than half of those were in IntervalArray and can be silenced be overriding IntervalArray._values_for_factorize to not issue the warning (on the theory that the warning is more useful for 3rd party authors than internal EAss). A bunch more are for MaskedArray and decimal.

The tests that would be annoying to fix are the arrow tests that already check for a PerformanceWarning.