Major Description
A user on the mailing list was sending document updates to a random node/core in his SolrCloud. Performance was not scaling anywhere close to what was expected. Basically, indexing performance was not scaling when adding shards and servers.
As soon as the user implemented a smart router that was aware of the cloud structure and could send to the proper shard leader, performance scaled exactly as expected. It's not Java code, so CloudSolrServer was not an option.
There will always be some overhead involved when sending update requests to the wrong shard replica, but hopefully something can be done about the performance hit.