Tables.columns should return a DataFrameColumns object

[oschulz]

Currently, we have

cols = Tables.columns(df::DataFrame) === df

Returning the DataFrame isself is not ideal, I think. A DataFrame is conceptually not really a collection of columns - for example, length(cols) is not defined, and you also can't map(f, cols) over the columns, nor can you iterate over the columns with [... for col in cols] or similar.

I think we should do

Tables.columns(df::DataFrame) = eachcol(df)

instead.

[quinnj]

The only things required of the return type of Tables.columns is that Tables.columnnames and Tables.getcolumn are defined, which DataFrame satisfies. I think what you're looking for is Tables.Columns(df), which provides the iteration/getindex/getproperty convenience methods you mentioned.

[bkamins]

Both DataFrameColumns and DataFrame support Tables.columns interface. Also if I understand things correctly Tables.columns does not promise an object that is "conceptually a collection of columns".

Can you please give more justification for the proposed change? The reason is that we would like to avoid being breaking till 1.0 release and what we have now conforms to the Tables.columns contract.

[oschulz]

Ah, right. Indeed, I should have used Tables.columntable(df) or Tables.Colums(df) - thanks!