Holding and Passing Custom Data between Custom Operators

[karthikeyann]

To pass Custom Data (libcudf dataframe https://docs.rapids.ai/api/libcudf/stable/classcudf_1_1table ) between Custom Operators, I would like design review for a proposed custom vector type CudfVector,

• CudfVector class derived from RowVector holds the libcudf GPU table as shared_ptr and passed around operators as RowVectorPtr
• Custom Cudf operators will dynamic_pointer_cast<CudfVector> and use it.
• New operators toVeloxVector and toCudfVector will convert data to and from velox.
• Only custom cudf operators are expected to get this custom datatype. Driver adapter or users will make sure to insert these conversion operators in plan fragment so that only custom operators will get this CudfVector

Is creating a new Vector type from RowVector an acceptable design?

[pedroerp]

Cc: @majetideepak @oerling @Yuhta @mbasmanova @xiaoxmeng

[pedroerp]

@karthikeyann thanks for starting this discussion. Like we discussed earlier, another option could be to specialize the Buffer class within a vector to contain a GPU memory location instead of host memory. What are the trade-offs between these two approaches? I suppose cudf::column has a buffer internally?

[Yuhta]

Yes inheritance would introduce strong coupling so any change on RowVector could break CudfVector in the future. Ideally if you can make it work with GPU memory backed buffers it would be the cleanest way.

[FelixYBW]

How close the cudf data format is to Arrow? Is it possible to reuse the arrow/velox conversion? In this way to the other Velox CPU operators, it's just a Arrow data source.