I'm really confused how this scales to high transaction rates. If the replica has to redo all the reads (which means talking to multiple nodes) before it can make a commit/abort decision for the transaction this could take tens of microseconds if all nodes are in the same datacenter (if serving from RAM). Since it also has to process transactions from the log in order, that seems like it would limit the transaction rate to tens of thousands of TPS? Forget about distributing a replica across data centers or having enough data that it may not be RAM resident.

Is this actually how it works or am I missing something important?