We have been seeing extended stalls during performance testing of MongoDB since the integration of WiredTiger with durable history. We need to investigate and undestand the root cause for those stalls. Then fix them!
It seems likely that the tests are getting into unbalanced replica set status, and at the end of the test a secondary is taking a long time to complete a checkpoint after the stable timestamp jumps forward a long way. Further analysis is required to understand whether that's actually the root cause, and if it is what we can do to resolve the problem.