The one thing to understand first
Aurora PostgreSQL keeps the PostgreSQL query engine but throws away its storage layer and replaces it with a purpose-built distributed service. The database tier never writes data pages to a disk. It writes only redo log records to a storage fleet, and that fleet is responsible for turning the log into pages. This one change — “the log is the database” — explains every Aurora feature.
The six-copy, three-AZ storage fleet
Aurora storage is a multi-tenant, log-structured, distributed system. Your data is sliced into small (10 GB) protection segments, and every segment is replicated six ways across three availability zones (two copies per AZ). Writes succeed once a quorum of four of six copies acknowledge the log; reads can be satisfied from a quorum of three of six. This design tolerates the loss of an entire AZ plus one more copy without losing write availability, and the loss of an entire AZ without losing read availability.
Because the storage nodes understand redo, they materialize data pages from the log in the background and continuously self-heal by comparing copies. The database tier is freed from checkpoints and full-page writes — the single biggest source of write amplification in classic PostgreSQL is simply gone.