Starting in 4.4, mirrored reads are enabled by default with a sampling rate of 0.10. If not disabled, the primary node mirrors supported reads (this includes find queries) to the secondary node's cache to "warm it up" to reduce performance impact of primary elections. Here the cache means "everything on RAM", which includes the plan cache. Thus even though the primary node executes a find query, the secondary's cache will "mirror" this read and have a plan for it in its plan cache. This results in the secondary having more plans in the cache than it executed. We disable this mirroring so the test passes in burn_in_tests.