Add sorting options to groupby

[bkamins]

Fixes #3251

[bkamins]

Thank you!

[jariji]

I think these can use only instead of getindex(_, 1).

[bkamins]

indeed it could. I initially had this test implemented on groups not on keys and using getindex works on both.

[jariji]

Having so many equivalent ways to specify sorting does seem a bit much? Not sure if it's worth doing anything about.

[bkamins]

The issue is that sort etc. provide that many ways to specify sort order, so we cannot do anything about it.
What is the rationale behind it:

• normally people will use the "global" settings like (rev=true,), which applies to all columns
• however there are cases when you want to specify sorting order per column, e.g. [order(:x, rev=true), :y], where you reverse :x but sort on :y in ascending order. Therefore order "per column" is needed.

In general - this complexity is needed when one has several columns.

[bkamins]

@jariji - if you would be willing actually improving https://dataframes.juliadata.org/stable/man/sorting/ section of the manual would be welcome. I planned to do it at some point, but maybe you would be willing to give it a shot and give a more in-depth coverage of all sorting options (this PR just inherits the complexity we allow for there).

[jariji]

Is there an advantage to sorting during the groupby versus sorting the groups afterwards?

[bkamins]

There is no convenient way to sort the groups afterwards AFAICT. To get a desired order you would need to sort the data frame you groupby before grouping, which is not always desired.

[jariji]

• Sorting before. I'm not sure what you mean it's not desired. Simply sort by the group columns before grouping.
• Sorting after. Introduce sortgroups(df). This is more flexible since it doesn't require regrouping in order to resort.

[bkamins]

I'm not sure what you mean it's not desired. Simply sort by the group columns before grouping.

It is expensive (in terms of time and memory)

Introduce sortgroups(df). This is more flexible since it doesn't require regrouping in order to resort.

Sorting while grouping will be faster and more convenient if user wants groups sorted (and most likely this is a most typical use case where user knows upfront how one wants groups to be sorted).

sortgroups(df) can be added as a separate function indeed. The question is how often re-sorting of groups is indeed needed (i.e. I would avoid adding a function that would be almost not used; the option to specify sorting order in groupby was added because users needed it).