[BREAKING] improve naming of ComposedFunction in aggregation

[bkamins]

Fixes change introduced by JuliaLang/julia#35980.

I assume we prefer function rather than ComposedFunction to be used in auto-generated column names. Also this is non-breaking this way.

@nalimilan - I have checked and I think that JuliaLang/julia#35980 does not break aggregation code for composed functions (we still should use a fast branch then), but maybe you will want to take a second look.

[nalimilan]

Thanks. Actually this is an opportunity to improve the printing of composed functions:

julia> repr(sum ∘ skipmissing)
"sum ∘ skipmissing"

Do you agree we'd better take advantage of this by adding a special method? One issue is that repr uses the full name, which isn't what we do know nor probably what we want (for conciseness):

julia> repr(Base.Libc.realloc ∘ sum)
"Base.Libc.realloc ∘ sum"

So I guess we have to reconstruct the composed name by hand.

[bkamins]

I was considering it but the problem is that things like: "sum ∘ skipmissing" are long, contain spaces, and many novices will not know how to key-in ∘. That is why I opted for printing what we currently do (i.e. :function). However, if we can find some concise way to print it I would be OK with it. A first idea would be to print e.g. sum_skipmissing - i.e. use _ and do not use ∘ at all. What do you think?

[nalimilan]

Ah, right, maybe using _ is more practical! We could also keep only the first function, and indicate that the other(s) is abbreviated in some way. The most common use case is f ∘ skipmissing and there we don't really care about skipmissing.

[bkamins]

so you would generate eg :sum_ for sum∘skipmissing?

[nalimilan]

I don't really know TBH. It's true that the trailing _ isn't very evocative. Maybe sum_skipmissing isn't so bad.

[bkamins]

OK - I will change it this way. Note that this will also affect multiply nested compositions (but this is rare, so it is probably not a problem).

[tkf]

I wonder if it's useful here if julia uses AnonymousFunction <: Function for anonymous functions?

[bkamins]

I think it would be useful so that we would not need to use the hack with #.

[bkamins]

OK to merge?
Also we should make a decision if we backport this (the change is breaking, but on the other hand without it DataFrames.jl behaves differently on different versions of Julia).

[nalimilan]

I think we'd better not backport this, as it could break some workflows. The inconsistency isn't ideal, but at least it will only hit people who upgrade to a newer Julia version.

[bkamins]

OK - I have removed backport. But this means that maybe we should consider making a release 0.22 when Julia 1.5 is out.

[bkamins]

Thank you!