DEPR: .select_dtypes accepts non-standard dtype references

[jreback]

xref #24541

DataFrame.select_dtypes() accepts strings including:

• datetimetz
• datetime64tz
• period (this is currently not-implemented).

These are non-standard names of dtypes, but not specific enough to the and should deprecated. An argument can be made that we should accept period to mean period[*] and datetime64[*, *] or something like this to select all sub-dtypes, but we con't really have any sytax to account for this.

[jreback]

cc @jbrockmendel @TomAugspurger

[TomAugspurger]

I'd support a way to select all the dtypes of parametrized dtype like datetimetz or period, much like we do for numeric. But what are you suggesting we deprecate?

[jreback]

the string datetime64tz and datetimetz are just made up things I think we need a format more specific (with a string), e.g. datetime64[ns, *] is more informative and compatible with what we have now.

[TomAugspurger]

I see, thanks. So taking `select_dtypes(np.number)` as example. You pass the class `np.number` to select every numeric-dtype column. Could users pass `select_dtypes(pd.PeriodDtype)` (the class) to select every period-type column? And possibly a string alias like `'period'`.

…

On Wed, Jan 2, 2019 at 11:06 AM Jeff Reback ***@***.***> wrote: the string datetime64tz and datetimetz are just made up things I think we need a format more specific (with a string), e.g. datetime64[ns, *] is more informative and compatible with what we have now. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#24558 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIkLXp5mN2vTH7Rx_UcKNpzGl-JNBks5u_OcOgaJpZM4Zm7w8> .

[jmg-duarte]

I just hit a use case where I need to be able to select period columns, what is the status on this issue?
I am available to help implement a solution!

[jreback]

this is an open issue and would welcome more specification and a PR

[jmg-duarte]

this is an open issue and would welcome more specification and a PR

What I meant was: is the idea to deprecate or implement?

For me it makes sense to use period instead of period[*] to refer to all period columns (likewise for the other types), although I don't care much for one or the other, I just need the feature

[jreback]

we can except both here, iow prob ok to joy depreciate anything as long as it's clear and what does what