Add opt-in Tables.isrow trait (similar to istable)?

[jrevels]

Imagine DataFrames wanted to support push!(df::DataFrame, row) for arbitrary row as long as it's a valid Tables.jl row.

It'd be nice to enable DataFrames to implement this like:

push!(df::DataFrame, row) = Tables.isrow(row) ? #= do it =# : throw(ArgumentError("hey that's not a row!")

(could have Tables.isrow(::AbstractRow) = true defined by default)

This would let DataFrames (and other packages) support non-AbstractRow rows in a manner that is still "safe" for values that are indeed row-like but probably intended to be scalars in a tabular data setting, so better errors can be thrown (i.e. avoid throwing a weird getcolumn MethodError if push!(df, "hello") is called)

[bkamins]

As a reference. Now, for safety reasons, in DataFrames.jl we only allow: Tuple, AbstractArray, AbstractDict, DataFrameRow and NamedTuple.

[rikhuijzer]

I also happen to stumble into this issue and also think that it could be nice. I think that it would be nice if Table.isrow(row::DataFrameRow) = true would be defined. This could help in StatsBase.predict, because that method now throws an ArgumentError for DataFrameRows. Instead, people have to manually construct a new DataFrame (or other Tables.istable object) from the DataFrameRow which is not very performant nor easy to use.

[bkamins]

Currently for predict this is required because this is what the contract for predict assumes (of course we could change this):

An object with new covariate values newX can be supplied, which should have the same type and structure as that used to fit model;

(the problem here is that the contract does not require newX to be a table - it could be anything - of course normally it is a table)