What?
JOINs in relational databases would be prohibitively computationally expensive
JOINs shouldn't involve computation and they're usually themselves in-memory lookups.
SQL might be shit for graph work but that has little to with graph databases. But graphs and topology are a different branch of maths than relational calculus.
As for transactional stuff: if it isn't ACID then it will break and you will lose data. Analytical processing can benefit from parallelism, just as it can live better with redundancy but the SparkSQL approach allows you to keep the API while playing with the storage.