Prior to version 3.6, the shard secondaries did not support filtering based on the active chunk metadata and were required to not have any filtering metadata cached. Because of this on every replication state change we were clearing it "just in case" - step up / step down.

With support for filtering in 3.6 this is no longer necessary and it also has negative performance impact due to the need to fetch the full metadata, so we should get rid of it.