A recent run of the ycsb.out_of_cache.95read5update test in the perf waterfall shows the dirty cache fill climb past 85% during a long checkpoint at the end of the test run. During this time WT is rarely conscripting application threads to help with eviction. This explains the high dirty cache, as the server is able to keep dirtying data even though WT isn't keeping up.

In contrast, a run from Jan 1 shows the cache behaving as expected – the dirty fill hovers around the 20% dirty eviction trigger, with application threads getting pulled in to help with eviction during checkpoints.

We should understand what changed – whether it's a WT change, a server change, or a test change, and decide whether and how to address it.