Return type of broadcast on GroupedDataFrame

[nalimilan]

Currently broadcasting on a GroupedDataFrame returns a Vector. This is inconsistent with map, which returns a GroupedDataFrame. Should we change this? I'd say yes:

• a Vector doesn't carry any information about the groups, making the result almost useless
• one can always use a comprehension to get a vector

If we agree to change this, we need to decide what to return exactly:

• a GroupedDataFrame: consistent with map, which makes sense since broadcast and map return the same kind of objects in general in Base
• a DataFrame: like combine, which may be more convenient since most operations are not supported on GroupedDataFrame

Maybe the solution is to return a GroupedDataFrame, but make that type behave more like a DataFrame (#1256). One issue is that a GroupedDataFrame doesn't make a lot of sense when each group contains a single row; so it depends on whether the most common use case for broadcast is to apply a function which returns multiple rows (like describe at #1539), or a single row.

[bkamins]

👍 for making broadcast and map consistent.

As of for what to do with GroupedDataFrame I am OK with whatever you would propose that is consistent and efficient as I guess you are understand the whole split-apply-combine infrastructure best (in the worst case the user can use combine after broadcast/map).

[pdeffebach]

I think a grouped data frame is nice. The big drawback is DataFrame to scalar or vector operations. Thankfully DataFrames can hold whatever kind of stuff they want.

[bkamins]

@pdeffebach

The big drawback is DataFrame to scalar or vector operations.

Can you expand your thought please so that I am sure what you mean exactly. Thank you.

[bkamins]

We now disallow broadcasting - and I would keep it that way. We can keep this issue open, but I would remove 1.0 milestone from it.