Use a consistent interface for Tables where input type cannot be converted

[bencottier]

I'm uncertain about this. There's a more general question of what operations on Tables we want to provide, vs. leaving to the user.

Currently, we have an apply! for tables that mutates the table columns in-place.

However this doesn't work when the input type can't be converted to the output type, e.g. from DateTime to Int for the hour-of-day transform. We could replace the column with the new type, but users should have a choice whether to do such a "strong" mutation.

Alternatively we could have an apply! that appends the new columns to the Table.

We could also have an apply that returns the transformed columns on their own (example), which gives users the most flexibility.

[nicoleepp]

I don't like

Alternatively we could have an apply! that appends the new columns to the Table.

just because apply! done on Arrays would not have the same behaviour and that's confusing

For that reason I think

We could also have an apply that returns the transformed columns on their own (example), which gives users the most flexibility.

is the best option, and let the user append the result to the table explicitly

[nicoleepp]

If we really wanted

Alternatively we could have an apply! that appends the new columns to the Table.

It should probably be a third function, like apply_append!(x::Table, ::Transfrom; newcolnames = [:a. :b], kwargs...)

or

apply!(x::Table, ::Transform; append=false, newcolnames=[:a, :b], kwargs...) but this last one might be confusing as sometimes it would mutate the cols that exist and sometimes append

[glennmoy]

is the best option, and let the user append the result to the table explicitly

It also gives them the flexibility to name it accordingly. Otherwise we'd have to come up with a naming convention in the package ☠️

It should probably be a third function, like apply_append!(x::Table, ::Transfrom; newcolnames = [:a. :b], kwargs...)

A slightly roundabout (but feasible) way is also

# for DataFrames
df.transformed = apply(df, ::Transform; kwargs...)

# for NamedTuples
x = apply(nt, ::Transform; kwargs...)
merge!(nt, (transformed=x,))

but if we had to make a single function call I'd prefer apply_append! as it's more explicit, can't be confused with normal apply! and doesn't rely on boolean kwarg.