Improve groupby sort

[bkamins]

Add an option to specify sort order in groupby.

[jkrumbiegel]

In sort this is done by passing lt = some_function which would be < by default I think. It's not very descriptive to match this, but could have the benefit of being the same as Base.

[jkrumbiegel]

There's also the by keyword which has the separate job of transforming the sorted values before sorting them with the function passed in lt. For natural sort order, as an example, you'd only need lt and not by, but in other circumstances by might be useful as well. It's just not good to have by as a keyword in groupby when it doesn't relate to the grouping.

My suggestions would be either, add keywords sort_lt and sort_by and mirror the sort keywords there. Or optionally pass a NamedTuple to sort keyword like sort = (; lt = func, by = ...).

[bkamins]

The challenge is the details. If lt and by should be "per column" or "per all grouping columns". Probably we should match what sort for AbstractDataFrame does.

[bkamins]

So, we would need to mimick the sortperm API. The changes would be:

• the sort kwarg apart from nothing, false, and true would accept a named tuple taking some or all of alg, lt, by, rev, order, with the same defaults as in sortperm; if such a named tuple were passed sorting would be performed.
• when passing grouping columns to groupby one could use order wrapper, e.g. groupby(df, [:x, order(:y, rev=true)]); if such order is passed then sort=true by default (passing nothing or false would error) (this is needed to allow for specifying sorting orders for grouping columns individually)

CC @nalimilan

[nalimilan]

Makes sense. That would just be a shorthand for groupby(sort(df, ...), cols) that avoids making a copy, right?

[bkamins]

This is what @jkrumbiegel wanted. Right?

[jkrumbiegel]

I think so, yes. It just seems natural to influence the sorting directly instead of splitting it out and doing sort = false on the groupby call.