DEPR: NDFrame.to_period, to_timestamp, tz_localize, tz_convert

[jbrockmendel]

These are methods that operate on the index/columns, not the series/frame values. We could deprecate in favor of e.g. obj.set_axis(obj.index.to_timestamp(), axis=0)

[topper-123]

#52023 was about adding a generic method for getting axes. Doing this + doing #52023 would IMO be much more flexible than the current situation and we wouldn’t need specific index methods on ndframes.

EDIT: changed "setting" to "getting".

[topper-123]

I'm +1 on deprecating those methods. i can never remember which index methods are mirrored in the ndframe, so I always end up just doing

df = df.copy()  # if I don't want to keep the old frame
df.index = df.index.tz_localize(tz)

IMO a set_axis method would be nice if it could be used in a pipe. For that to work, the first argument should be either an indexlike or a callable, so something like piped_df.set_index(lambda x: x.tz_localize(tz), axis="index") and piped_df.set_index(pd.Index(...), axis="index") could both work.

We have a lot of index methods and only some will feature in the axis_ops namespace and it's not intuitive why those have been selected and others not. so it's same problem as beforehand for me. What about adding a set_index method and skipping axis_ops?

[Dr-Irv]

We had a discussion related to this PR during the dev meeting on 4/26/23. Methods like NDFrame.first(), NDFrame.last(), NDFrame.between_time(), and NDFrame.filter() are really creating an NDFrame based on values in the index, and that it might be better to move those methods to be on Index, and then use .loc on the resulting index to get the result (which is most of these implementations do anyway).

I did a PR for filter() to create a new Index.filter() method that would give the index, but it didn't go anywhere.
#51370 The discussion we had was maybe we should be moving more of these methods out of NDFrame and maybe into the new axis_ops proposed here.

[jorisvandenbossche]

I am fine with deprecating to_period and to_timestamp (and I don't think we necessarily need a new alternative like axis_ops for those, I don't think those are used that often), but personally I would prefer to keep tz_localize and tz_convert