To refresh or remove session records from the logical sessions collection, the sharded logical sessions cache (i.e. on a mongos or shard server) uses lambdas that run bulk writes through the ClusterWriter (here and here). These lambdas return the status from each write's response, but when the lambda is invoked, the return value is ignored, so if either write fails, the refresh will not notice and continue running, which can lead to incorrectly killing the sessions of recently established cursors.

It looks like this was introduced by SERVER-42508, so it should only affect 4.4 and the current master.