The collection defragmentation algorithm and the cluster chunks resize policy generate a continuous stream of mergeChunks and splitChunk commands issued by one of the balancer thread: each operation, when committed, triggers a refresh of the metadata caches on the secondary nodes of the affected shards (details in SERVER-62698), introducing some latency to concurrent CRUD operations.
In order to limit the performance impact, the rate of mergeChunks and splitChunk requests needs to be throttled.