PAPERS-15-008 · Paper — Amazon Aurora — The Log is the Database

The paper

Amazon Aurora — The Log is the Database, authored by Amazon engineers in 2018 and presented at the VLDB conference, is considered a landmark because it reimagined relational databases for the cloud and showed dramatic performance gains.

Before this paper

Before Aurora, cloud databases tightly coupled compute and storage and replicated full data pages across nodes. Shipping whole pages was slow and consumed a lot of bandwidth. This design limited scalability and made crash recovery expensive. The field needed a way to reduce network traffic while keeping durability.

The key idea

Aurora’s key insight is to treat the redo log itself as the database. By moving the storage engine into a separate, distributed, log‑structured service, only compact log records are sent over the network instead of full pages. This decouples compute from storage and makes the log the single source of truth.

How it works

When a database instance processes a transaction, it writes redo log records locally and immediately streams them to Aurora’s purpose‑built storage layer. The storage layer stores the log redundantly and replays it on demand to reconstruct data pages for reads. Data is replicated six ways across three availability zones. Writes succeed once four of the six copies acknowledge, and reads succeed with any three copies, allowing the system to lose an entire zone plus one extra copy without losing durability or availability. Because only small log entries travel across the network, write amplification and latency drop sharply, and storage can scale independently of compute.

See it — push the log to storage

flowchart TD
  A["Database engine"] --> B["Send only the<br/>redo log"]
  B --> C["Storage layer<br/>replays it"]
  C --> D["6-way quorum<br/>across 3 zones"]

What they showed

The paper demonstrated that Aurora achieved substantially higher throughput than running the same engines in a traditional setup, with fast crash recovery and strong durability, all while remaining compatible with MySQL and PostgreSQL.

What it changed

Aurora popularized the disaggregated, log‑as‑database architecture, inspiring many later managed cloud databases to separate compute and storage. This shift enabled more elastic scaling, lower latency, and higher availability across the industry. Next: the Evidence library.

The one-line takeaway

Treat the log as the database to decouple compute and storage and dramatically improve cloud database performance.

Read it / meet it

Reading the Aurora paper reveals how a simple change in architecture can unlock big gains, and you’ll see its legacy in today’s managed services like Amazon RDS, Google Cloud Spanner, and Azure Database for PostgreSQL.