The problems that Jepsen found were centered around the "transactions" feature that Cassandra added. We don't use these and don't need them since we don't need 100% consistency and prefer availability (for example we read at quorum to trigger read repair, but downgrade to single node reads if we need to).
Also ScyllaDB is a new product and it would be crazy to start off with it. We plan to run a long-term double write experiment before we are comfortable with using it as a primary data store.
The Jepsen tests were not completely centered around transactions. It also had to do with data loss when replicas go down and pure "last-write-wins" approach. For those wanting more info around this the original post is here:
Also ScyllaDB is a new product and it would be crazy to start off with it. We plan to run a long-term double write experiment before we are comfortable with using it as a primary data store.