In WT-2869 and WT-2954 we worked through a case where appending to a page could run extremely slowly because once the page grew to the maximum size permitted in memory, every access of the page would try (and fail) forced eviction 10 times.

The workload in those tickets and some internal tests are now showing that there has been a regression and the same issue has been reintroduced. As well as fixing the regression, we should automate the workload in the linked tickets and set criteria so that it fails when read throughput cannot be sustained at the expected level.