Graviton5 R9g on Aurora and RDS: should you move?
AWS launched Graviton5 R9g database instances for Aurora and RDS. Here is what changed, who gains, and how to test a switch safely this week.
On November 12, 2026, AWS announced Graviton5-based R9g instances for Amazon Aurora (PostgreSQL and MySQL compatible) and Amazon RDS for PostgreSQL, MySQL and MariaDB. AWS claims up to 30% better performance than Graviton4 instances of the same size, measured on Aurora. If your database is a large line on the AWS bill, this is worth a test. It is not a reason to change production this week without one. This post covers what changed, who should care, and what to check first.
What did AWS announce?
According to the announcement, posted November 12, 2026, R9g instances are generally available. They run on AWS Graviton5 processors and the AWS Nitro System. AWS lists these facts (as of November 2026):
- Nine sizes, from large to 48xlarge, with up to 192 vCPUs.
- A memory-to-vCPU ratio of 8:1.
- Up to 100 Gbps of network bandwidth and up to 72 Gbps of bandwidth to Amazon EBS.
- Availability in US East (N. Virginia), US East (Ohio), US West (Oregon) and Europe (Frankfurt).
The announcement does not list Canadian regions. If your data must stay in Canada, you cannot use R9g there yet, based on that list. Confirm region support in the RDS console before you plan anything, since lists change.
What actually changed compared with R8g?
R8g is a Graviton4 memory-optimized class listed in AWS's documentation, and R9g is the Graviton5 one. AWS compares R9g with Graviton4-based instances of equivalent size, measured on Aurora:
- Up to 30% better performance.
- Up to 22% better price-performance at on-demand pricing.
AWS adds that results depend on the engine, the version and the workload. That sentence matters. "Up to" is a ceiling measured by the vendor, not a forecast for your queries. A database that is bound by storage I/O, locks or one slow query may gain little from faster cores.
Source: AWS What's New, November 12, 2026. A vendor ceiling, not a measurement of your workload.
Who should care, and who can ignore it?
Pay attention if you:
- Run Aurora or RDS on large or many instances, so a single-digit percentage matters in dollars.
- Hit CPU limits on your writer or readers before you hit storage or connection limits.
- Are already on Graviton and keep your instance class current.
- Run in one of the four launch regions.
You can mostly ignore it for now if you:
- Run small instances or burstable classes. R9g starts at the large size, and the announcement does not describe burstable options.
- Are CPU-idle most of the day. A faster processor will not lower a bill for headroom you do not use.
- Are bound by a region outside the four listed.
- Run an engine or version not on the supported list (next section).
What limits should you check first?
AWS's instance class table for Aurora shows that R9g support is tied to specific engine versions, and the list has gaps. For example, as of November 2026, Aurora PostgreSQL 17.4 through 17.7 and 17.9 through 17.11 are listed, and 17.8 is not. The Aurora MySQL list includes 3.08 through 3.14 in specific minors, and 8.4.7, 8.4.8 and 8.4.10. The older R8g class, by contrast, is listed as "and higher." Read the table for your exact version rather than assuming support.
The announcement does not say which RDS engine versions are supported. It points to the RDS instance class documentation for that, so check it before you plan an RDS move. For example, that page lists RDS for MySQL support as "8.4.5 and higher" only, so an RDS MySQL 8.0 database would need a version upgrade first (as of November 2026).
Cost is the other limit. The announcement does not give prices; it points to the Aurora and RDS pricing pages. Compare R9g and your current class on the Aurora pricing or RDS pricing page, in your region, for your engine. If you hold reserved instances or other commitments, read your agreement to see whether they cover a new generation, because the announcement does not say.
How do you try it safely this week?
AWS's Aurora modification guide says to test changes on a test cluster before you change production. It also says that changing a DB instance class causes an outage for that instance. Other instances in the cluster have no downtime from that change. So plan the change as a controlled event, not a quiet setting.
- 1Check version and region
- 2Clone or restore to test
- 3Replay real load
- 4Compare cost and latency
- 5Roll one reader
- 6Fail over the writer
An approach, not an AWS procedure. Failover behavior varies by setup, so rehearse it.
Here is the order we would follow:
- Confirm eligibility. Your region, engine and exact version must be on AWS's R9g list.
- Build a test copy that matches production. On Aurora, a clone or restore works; on RDS, restore a snapshot. Same engine version, parameter groups and data volume, as far as you can.
- Replay real traffic. A synthetic benchmark tells you what the vendor already told you. Your own query mix tells you what you will get.
- Compare three numbers. CPU use at the same load, p95 latency, and cost per hour for the class. The price-performance claim is only useful if all three line up.
- Switch one reader first in production, off-peak, and watch it for a few days.
- Move the writer last, in a maintenance window, after you rehearse the failover with your application's connection handling.
Keep the old class as your way back. Because the class change is per instance, you can return one instance at a time if something looks wrong. Changing back is also a class change, so expect the same outage on that instance.
What would we check before adopting it?
We start with whether the database is actually CPU-bound. If CPU is low and latency comes from I/O, locks or a missing index, a new instance class is the wrong lever, and a query fix costs less. We would also check these:
- Right-sizing first. A move to R9g on an oversized instance locks in waste. Shrink first if the data supports it, then change generation.
- Commitments. Do not buy a new long commitment on a class you have not tested.
- Connection behavior. Confirm your application reconnects cleanly after a failover.
- Rollback. Write down how you return to the old class and who decides.
For a SaaS team with a database bill that keeps climbing, a measured test over a few days costs little. A blind switch can cost an outage and still not save money. If you want a second pair of eyes on the numbers, our services page describes how we work on cloud cost and architecture, and fractional CTO support covers the wider decisions.
General information, not financial or technical advice. Pricing, regions and supported versions change, so confirm them against AWS's current documentation before you decide.