Redshift RG instances debut in Tokyo

Amazon Redshift just got a new hardware generation — and this one is really about collapsing two analytics worlds into one. Warehouses and data lakes have been drifting together for years, but the plumbing has stayed awkward. You kept hot, structured data in Redshift, colder or bigger data in S3, and then paid an extra coordination tax to query across both. On May 12, AWS introduced Redshift RG instances, powered by Graviton, with a built-in data lake query engine. Tokyo appears to be one of the regions where customers can already launch them. ### What is an RG instance? RG is a new Redshift provisioned node family. It sits next to — and is clearly meant to supersede — RA3 for a lot of workloads. AWS is pitching RG as the default choice on the Redshift pricing page, which is a strong tell about where the product line is headed. The basic idea is simple: keep Redshift’s managed-storage model, swap in AWS Graviton CPUs, and add a query engine that can hit S3-resident open-table data directly from the cluster. (aws.amazon.com) ### Why does the S3 part matter? Because this is the real story, not just the chip swap. Redshift has long been able to reach into data lakes, but often through Redshift Spectrum’s separate scanning layer. RG changes that by pulling data-lake execution into the cluster’s own engine for Apache Iceberg and Parquet. Basically, the warehouse and the lake stop feeling like two systems stitched together at query time. (aws.amazon.com) ### What are the actual performance claims? AWS is making three big ones. First, standard warehouse workloads can run up to 2.2x as fast as RA3. Second, Apache Iceberg workloads can run up to 2.4x as fast. Third, Parquet workloads can run up to 1.5x as fast. On price, AWS says RG comes in at 30% lower cost per vCPU than RA3. Those are vendor benchmarks, so real-world results will vary, but the direction is clear — cheaper compute plus faster lakehouse-style queries. (aws.amazon.com) ### Why launch this in Tokyo? Because region availability changes whether a launch is real for customers or just a roadmap slide. A Tokyo-based AWS partner has already shown RG clusters running in ap-northeast-1, including an `rg.xlarge` cluster created successfully there. That matters for APAC teams with data residency needs, local latency concerns, or procurement rules that make “available somewhere else” basically useless. (aws.amazon.com) ### Is this just about lower cost? Not really. Cost is the bait, but operational simplicity is the bigger win. If your analysts, dashboards, ETL jobs, and newer AI-style agents all hit the same estate, every extra engine boundary becomes latency, management overhead, and surprise billing. RG’s pitch is that one engine can cover warehouse tables and lake data together — with fewer moving parts and, in many cases, no separate Spectrum per-terabyte scan charge on those integrated lake queries. (dev.classmethod.jp) ### What’s the catch? A few things. Right now the family looks narrower than RA3. The Tokyo hands-on writeup shows `rg.xlarge` and `rg.4xlarge`, no single-node mode, and a minimum of two nodes. It also notes there is no RG equivalent yet for `ra3.16xlarge`. So this is not a full one-for-one replacement across every cluster shape today. (aws.amazon.com) ### Where does SeekStorm fit in? Only as background texture. SeekStorm’s recent work around vector and hardware-specific optimization does reinforce the broader Graviton-and-ARM momentum in analytics infrastructure, but it is not part of the AWS launch itself. The core news here is Redshift changing its own execution model, not a partner ecosystem announcement. (dev.classmethod.jp) ### Bottom line This looks like one of those infrastructure releases that sounds incremental but changes buying decisions fast. If Redshift customers in Tokyo can get materially better Iceberg-on-S3 performance, lower per-vCPU pricing, and fewer cross-engine headaches in the same move, RG is not just a new instance type — it is AWS trying to redefine what a Redshift cluster is for. (aws.amazon.com)