The clusterwide_ops_with_add_remove_shards.js verifies that those operations that must be run on all shards do not crash when adding or removing shards. As the CheckMetadataConsistencyInBackground is run concurrently with that jstest, it is possible for the mongos to refresh its ShardRegistry after choosing to target one shard, throwing ShardNotFound error.

One possible solution is to handle ShardNotFound error in the same way as CheckMetadataConsistencyInBackground does for retryable errors.