Ok. From skimming through the paper, it seems that for a given transaction, only the replica sets of the shards touched by this transaction need to participate.

Ok. From skimming through the paper, it seems that for a given transaction, only the replica sets of the shards touched by this transaction need to participate. Thus for two transactions touching disjoint sets of shards, they can run completely in parallel. Hence the scalability. Is that right?