Abstracting Spark API

I would like to understand if anyone in the community has built services that effectively perform calculations on an underlying distributed datastore (Hadoop/Parquet files) using Spark but at the same time been able abstract out the Spark API functionality. This may not seem counter intuitive given the compute paradigm employed in spark i.e send the compute to the data but our architecture and management teams are keen on building vendor/technology independence upfront. Are there any design patterns that can help with this abractions? The typical presentation, service, DAO tier found in many applications to separate your business logic doesn't seem relevant here I suspect.

I would we very interested if anyone else has done this or even been given the same challenge.

Thanks