After a chat with Ops we decided to move away from RAID0 arrays for the Cassandra instances partitions and to replace them with RAID10 ones.
New architecture depicted in https://wikitech.wikimedia.org/wiki/User:Elukey/Ops/AQS_Settings
Since we have already loaded 4 months of data, we'll attempt to re-image one host at the time giving time to Cassandra to rebuild 2 instances at the time streaming data from the other ones.