Triffids

It depends. There are Cloudera kudu engine, it quite easy can handle millions atomic inserts per second. But it can't support true acid transaction. On other hand SQL also can't use true ACID in high load system. Serialazible transactions replaced by Read committed transactions, "banking day", queues + background process, etc

Generally good idea, analytical queries will be faster than on rdbms. You problem in strange tools and format. Hive = map reduce, JSON = scan for all data. You need to replace map-reduce (Hive) to Impala or Spark 2.x and JSON to columnar format: Parquet or ORC.