PERF: Investigate caching on PeriodArray

[TomAugspurger]

Start with benchmarks

1. ops on PeriodArray and PeriodIndex
2. ...

If we determine that caching attributes like hasnans and isna does help then we can investigate caching on PeriodArray.

[jbrockmendel]

Could also be useful for NumpyBackedExtensionArray

[TomAugspurger]

We'll need to be especially careful here. The point of PandasArray (formerly NumPyBackedExtensionArray) is to be a zero-copy view on some other ndarray, which makes failing to invalidate a cache more likely.

…

On Tue, Dec 18, 2018 at 10:26 PM jbrockmendel ***@***.***> wrote: Could also be useful for NumpyBackedExtensionArray — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#23359 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIjBMrM-zH4Fg3xguEb_RZgyHqydRks5u6b_3gaJpZM4X8PLp> .

[jbrockmendel]

Been thinking about this recently, I think the view trouble is going to make caching on DTA/TDA/PA infeasible

[TomAugspurger]

Is the trouble PeriodArray taking a view on another array, or another array taking a view on the PeriodArray?

[jbrockmendel]

its one PeriodArray taking a view on another, so if the first does a setitem (that presumably invalidates the cache), the cache would need to be invalidated on the other

(this is also why im increasingly convinced that DTA/TDA need to have freq removed)

[TomAugspurger]

Gotcha. In theory, PeriodArray.__init__ could create a weakref in both directions (view to orig and orig to view) and __setitem__ would invalidate the caches of both. But that feels increasingly fragile.

[jbrockmendel]

Closing as not worth the trouble