The current Cassandra source sink is tightly coupled to the Datastax Object Mapper. This requires users of the sink to provide a POJO describing the table in Cassandra. Although the POJO is a easy and powerful way to describe the table it does requires users to always recompile the pipeline for each change in the table definition.
I suggest adding a abstraction layer that allows users to inject their own mapper implementation. One example use would be a mapper that works with a generic Row implementation rather than a compile-time POJO.
This abstraction layer could be implemented by simply supplying a MapperFactory through the Sink/Source builder.