It is about dividing the tasks, one per graph node per block is not the right solution - every transaction involves two nodes by user, and multiple nodes by token. Maybe one per payer. Synchronisation with massive parallelisation requires prioritisation of propagation, ensuring frequent transactions to be clustered, but whenever possible, spread across many nodes, by frequency. Geographical aggregation and association aggregation form transient maps that show how to keep sync without total convergence of blockchain state. So long as the nodes always see the current state of the parts they work with (to some degree of immediately visible provenance), it doesn't have to fully converge, or be permanently stored.

There is many approaches to this, i am very interested to see what people turn up.