In testing https://gerrit.cloudera.org/#/c/9758/ I noticed that TPC-DS Q67A at SF 300 takes 49525M to run without spilling. This is an order of magnitude larger than any other query. The query is too large for the cluster to also stress the cluster concurrency-wise.
The stress test has an option that in theory is supposed to catch this
but it's broken. Fix/improve the option.