I feel that the challenges will be far more due to the limitations of the workflow & Databricks. For example- How to install the libraries dynamically to the cluster etc.

And also this type of framework should not be called as ETL framework. Because it will not work for all the use cases. 
At best you can separate it out and call it Orchestration Framework because you can not drive the transformations through the metadata. You can only orchestrate your notebooks through a framework (including few key details like table names and stuff.)