For me, it's two main reasons: 1) Generating and storing massive amounts of pre-computed permutations of the same data based on how I'll read it with easy linear growth by adding servers and 2) Transparent multi-master reads/writes providing easy HA failover. Note, not all NoSQL options offer this. If existing free relational DBs had these options easy and out of the box, I would use them instead.

I just need a ton of data, often redundant, that I never delete, that supports distributed writes and reads (ACID not required). The things I use Cassandra for would require huge Postgres servers and custom sharding/partitioning setups and Citus HA and all sorts of other things.