> The key insight of our solution is decoupling a durable transaction into three > fully asynchronous steps. (1) Perform: execute the transaction on a shadow > memory, and produce a redo log for the transaction. (2) Persist: flush redo > logs of transactions to persistent memory in an atomic manner. (3) Reproduce: > modify original data in persistent memory according to the persisted redo logs. > It is essential that we never directly write back dirty data from shadow memory > to persistent memory – all the updates are realized via the logs.
It is exactly the same Amazon did for its Aurora! Using this decoupling, Aurora provides great replication and failover (by use of replicated and versioned storage both for logs and for data pages).