On sharded clusters, some aggregation stages such as $lookup, $graphLookup, $out and $merge may choose to split the pipeline and designate a merging shard. The router chooses the merging shard according to its cached routing table (for the inner collection in case of $lookup/$graphLookup, and the output collection in the case of $out/$merge). This routing table can potentially be stale – if it is, the query plan will perform sub-optimally, but results will be correct.
This ticket is to prevent scenarios where a router stays stale indefinitely and keeps choosing sub-optimal plans.