The first time a config server becomes primary it builds the necessary indexes on the config database and initializes the config.version collection, which includes the cluster ID. It then immediately loads the cluster ID into memory. If the config.version document rolls back, however, we don't clear the cached in-memory cluster ID, so when a new config.version document is written, with a new cluster ID, we won't ever realize that its changed. This means that when we add a shard, we'll provide it the wrong clusterID in its shardIdentity document.