While it shows large improvement, the long-running (2 hour) large-data Jenkins test for read-only is showing wildly varying throughput. The throughput numbers used to be very stable and are now swinging by 200% since the merge of WT-4433 for dynamic compression.

This is only on the read-only test. The earlier, also long-running, tests for populate and read/update mixes do not show this variation. The plot is here: http://build.wiredtiger.com:8080/job/wiredtiger-perf-long/plot/

I have attached the current version of the plot that illustrates this problem in case the link ages out of the pre-merge information.