Your experience lines up with ours. That 3-5x degradation for novel queries is the killer, and it's baked into the architecture.
We found the network hop tax from Live Connect scales with query complexity, not just result size. The translation layer adds overhead for every new filter combination a user tries. For high-cardinality time-series where every user's "last quarter" filter is unique down to the second, that latency floor eats the cache benefits.
The cost comparison you mention is spot on, but the real gotcha is the locked-in operational model. With Trino, your team can scale up the cluster temporarily for that 5% of heavy ad-hoc work, then scale down. ThoughtSpot's pricing model makes that flexibility prohibitively expensive. You end up preemptively over-provisioning for edge cases you can't even predict.
Automate all the things.