Allow `dup_axis` within `scale_x_date`

[brianwdavis]

---

A question on SO led to wondering why duplicate axes aren't usable with scale_x_date or *_datetime. I worked up fairly hacky solutions by converting to numerics and working back from date_breaks, but it should "just work". While you're at it, I'm sure it would be nice to be able to have transformed sec_axis: e.g. hours on the x-bottom and days on the x-top.

See reprexes at: https://stackoverflow.com/a/45373735/3330437

[hadley]

@thomasp85 any chance you have time to take a look?

[thomasp85]

When do you expect the next release? I can’t promise anything this week

[hadley]

I'll probably start the release process in 2-3 weeks.

[thomasp85]

I'll give it a shot

[thomasp85]

Do you have any feelings regarding the transformations allowed with dual axis in the current implementation? To me it seems dual date axes is mostly for decorative purposes, though someone might conceivably want to plot a timeserie with both Chinese and Gregorian calendar..?

[hadley]

Yeah, it's not clear that this needs to support transformation, but maybe just different breaks/labels? (i.e. have years on top and months on bottom)

[thomasp85]

That’s my feeling too, but is also the solution resulting in the worst API. It would require an additional set of sec_axis/dup_axis functions that blocks the transformation argument..?

[hadley]

Hmmm, but it probably needs a different sec_axis anyway for date_breaks() etc?

[thomasp85]

I suggest we postpone it to after this release. I’m still interested in doing it but I don’t want to butcher the API to get it done, so I’ll need some time to think it through...

[hadley]

Ok, np.

[kevinwolz]

Another feature related to this is the ability to use two different columns/datasets from which to base the time axes. For example, I have a dataset with "actual time", which I would like to have label the bottom x-axis in date format, and "time since event", which I would like to have label the top x-axis in hours format.

This would be very useful for me, but I also have a feeling that it would be useful in many time series analyses. Not sure what the best approach would be to accomplish this, though.