> It is absolutely without a doubt 100% not diverse enough.
sure, nothing is 100%. But 90% may be good enough.
> JSON querying and the like. Nearly every organisation has something in that space now, and TPC-H/TPC-DS don't touch it at all.
unnesting json into relational data is some trivial op, and then you come back into tpc-h/ds realm.
> unnesting json into relational data is some trivial op
I mean, everything is representative if you just phrase the bits you don't want to do as "some trivial op". By this logic, Clickbench is a great and totally representative benchmark because joins are also trivial ops.
Billions of dollars are spent processing semi-structured data. It's a critical part of most analytical pipelines. You can't just ignore it and pretend like your benchmark is representative.