Much of the world runs on relational databases. If you want to run one in AWS, there are two popular choices: Amazon RDS and Amazon Aurora.

These can look quite similar. They’re both managed services, where you pay Amazon to manage and administer your database. They both let you spin up databases with a few clicks in the console. They both scale to dizzying heights, with terabytes of storage per database.

These similarities can make it hard to tell them apart. 

Which begs the question: When should you use RDS and when should you use Aurora?

Four approaches to database management

If you run a database in your own data center, you’ve probably hired database administrators. Your DBAs run databases on servers that you own. They install software, apply maintenance and security patches, make regular backups, and so on. It’s a labor-intensive process—but without the cloud, it’s how you do things.

As a first step toward the cloud, you can run a database on EC2. This is like running a database in your data center but with EC2 instances instead of servers. This is less and less common, but it’s useful if you need very fine-grained control over your database.

Amazon created RDS to reduce the management overhead of running a database in EC2. You click a few buttons, and you get a database ready to store data. Amazon handles all the fiddly bits—provisioning instances, replication and backups, maintenance and updates. It runs a variety of database engines—including MySQL, PostgreSQL, MariaDB, and Oracle—so it works with your existing application code.

If you have databases you want to migrate to the cloud, RDS is a great start.

(If you are migrating, talk to your AWS account manager first. Amazon wants to help you move your on-prem databases into RDS, and your account manager may be able to provide credits or professional services to help you along the way. They also have the Database Migration Service, an underrated tool that does exactly what it says.)

But RDS still looks like a database running in a data center. What if you didn’t have to imitate an existing architecture?

Amazon created Aurora to be a cloud-first database. Externally, it behaves like any other RDS database. It’s API-compatible with MySQL and PostgreSQL, and it’s meant to be a drop-in replacement. Amazon still handles all the fiddly work of managing the database—but under the hood, it’s quite different.

Rather than running the entire database on a fleet of EC2 instances, Aurora splits the compute and storage into different pieces. Storage is handled by a custom data layer, designed to take advantage of Amazon’s cloud infrastructure. This “secret sauce” has a number of benefits, so let’s dive in a bit deeper.

Aurora’s cloud-first architecture

Although Aurora is closed-source, Amazon is pretty open about how it works. You can find a variety of re:Invent sessions, documentation pages, and white papers about the Aurora architecture—if you’re interested, do read further! Here, I’m just going to present a high-level overview.

In a traditional database cluster, you have one or more nodes (servers or EC2 instances): a read-write primary/writer (W), and read-only replicas (R). If you write to the primary node, that write gets synchronously replicated to the replica nodes. Different database engines use different consensus protocols for replication to ensure all the nodes have a consistent view of the data.

Each node is responsible for both compute and storage, so the entire database is contained in these nodes. If you want more durability or scale, you need to add more replicas—and that comes at a cost. More replicas means more replication traffic, and at some point network I/O becomes a bottleneck. It also takes time to modify the cluster, because new nodes have to replicate all the existing data before they can serve queries. To keep replication manageable, RDS limits you to five replicas.

In an Aurora cluster, there are different nodes for compute and storage.

Data is stored in a shared “cluster volume”, which spans six storage nodes and three availability zones. This gives you multi-AZ resilience, and it uses the same “pay for what you use” model as S3. Rather than provisioning storage upfront, the volume grows or shrinks to match your data, up to 128TB—and your bill grows or shrinks to match.

Queries are handled by compute nodes, which all talk to the same cluster volume. You have the same primary/replica split, but replication is handled entirely within the storage nodes. There’s no synchronous replication between compute nodes, and they don’t hold any permanent state.

This architecture has a number of benefits:

Aurora scales faster because it can add new read replicas quickly. Because replicas all use the shared storage volume, a new replica can serve queries almost immediately. It doesn’t have to wait to replicate data from the other nodes.

Aurora scales further because extra compute nodes are cheap. Aurora does some asynchronous cache replication between nodes, but nothing synchronous. This reduces the inter-node I/O, which means Aurora can have more replicas; where RDS allows just five, Aurora allows 15.

Aurora is more durable. You always have multiple copies of your data. Every Aurora cluster has six storage nodes, spread across three AZs, even if you have just one compute node. Even if an Availability Zone suffered a spontaneous existence failure, your data would be safe. In RDS, you have to max out your read replicas for this level of durability.

Aurora is more resilient. It has fast recovery from failures. If a compute node crashes, Aurora can recover quickly. It can start new read replicas with minimal lag, and if the writer fails, another replica can be promoted to take over without waiting for the other nodes to reach consensus. All the shared state is in the data nodes, so failed nodes can be replaced almost immediately.

These are all desirable properties. But in addition to this, Aurora’s architecture enables another approach that’s worth discussion separately.

Aurora Serverless

Aurora’s storage design means you don’t have to provision storage in advance; it scales to meet demand. What if the same was true for compute?

Aurora Serverless lets you run Aurora without having to guess how many compute nodes you need. It automatically starts and stops nodes to match the needs of your application. It scales up to meet a spike in demand and scales down when things are quiet. The data remains in the shared storage volume, independent of any scaling.

This is particularly useful if you have spiky, intermittent, or unpredictable workloads. You only pay for the compute capacity you use, so you could enjoy big savings if you have long idle periods. (But good luck predicting your spend in advance!)

If you need sustained workloads, look at using regular Aurora or RDS instead.

What are the limitations of Aurora?

Aurora sounds great: more scalable, more durable, more resilient—who wouldn’t want those things? But before you jump in, here are a few reasons you might choose RDS over Aurora:

You use a database engine that Aurora doesn’t support. RDS supports more database engines and features than Aurora:

  • RDS supports five database engines; Aurora just two. If you need MariaDB, Oracle, or Microsoft SQL Server, RDS is your only choice.
  • Although Aurora is API compatible with MySQL and PostgreSQL, it’s not always the newest version, and it doesn’t always have the same features.
  • Aurora only uses the InnoDB storage engine. You can’t use other storage engines.

You want to use the AWS Free Tier. You can run an RDS database for a year in the free tier; there’s no free version of Aurora.

You want predictable database costs. RDS has the same pricing model as EC2; there’s a fixed per-hour cost for every instance you run. Aurora pricing is harder to predict. In addition to the hourly instance cost, you pay per million I/O operations. How many will you use? It’s hard to tell in advance.

Your workloads aren’t a good fit for Aurora. Amazon marketing tout statistics like “5X the throughput of standard MySQL” and “3X the throughput of standard PostgreSQL.” But database performance can’t be reduced to a single number. Aurora will be significantly faster for some workloads, sure. But it might be slower for others. If you’ve already fine-tuned your queries for a particular database engine, or Aurora seems slow, you might want to benchmark against regular RDS.

Should I use RDS or Aurora?

For many organizations, RDS is the “safe” choice. It’s a longstanding, reliable service that looks like the databases they already have—with predictable pricing and similar alternatives as other providers.

Aurora’s unique architecture gives you more durability, scale and resilience. And for many workloads, it’s cheaper and faster than running the equivalent RDS database. It’s what I’d choose, if I was building a new application today—but it may not be right for you.

Bottom line? Aurora can be a big win over RDS, but it’s not a slam dunk. There’s no “right” way to run a database, which is why Amazon gives us the choice.

Need some help choosing a relational database in AWS? Give us a holler.