If the config server primary steps down right after removing the config.shards doc for the shard but before responding with "state": "completed", the mongos would retry the _configsvrRemoveShard command against the new config server primary, which would not find the removed shard in its ShardRegistry if it has done a ShardRegistry reload after the config.shards doc for the shard was removed. This would cause the command to fail with ShardNotFound.