We have an organization full of analysts who know SQL very well and write complex queries. Moving them all to a mapreduce query language proprietary to mongo, riak, or cassandra (and then being locked into it) would have been much more painful. Postgres has worked out very well, and I had confidence in it from prior experience which influenced the decision. With proper partioning in the schema and a json type we have the best of both worlds (SQL+NoSQL) as far as this system needs.

Or alternately: Why wasn't the JSON parsed into normal form if the intent was to search & filter across that data.

To get the same flexibility you'd want from a nosql data store. Without giving up the simpler, more open access methods. The schema is partly predictable (and indexed or fielded on those parts) and partly unpredictable.

why not add the partly predicted keys as columns ?

also jsonb will require a little more storage

is 8tb in 1 tablespace/db ?