Databases

AI agents asking questions in natural language apparently issue a lot more queries than your average SQL jockey

AWS says it has boosted Redshift performance by bringing the data warehouse to its Graviton-powered instances,  positioning it to handle AI agent workloads and fend off rivals in the analytics market.

Redshift's new RG instances, powered by AWS Graviton processors, accelerate new query workloads by up to seven times. AWS claims the instances are up to 2.2x faster than the RA3 family, which it introduced in 2019, at 30 percent lower cost per vCPU. 

The updated query engine also lets users run SQL analytics across data warehouses and data lakes from a single engine, delivering up to 2.4x the performance of RA3 for Apache Iceberg and up to 1.5x for Apache Parquet.

Amazon Redshift RG instances are available in AWS Regions including US East, US West, Asia Pacific, Canada (Central), Europe (Frankfurt, Ireland, Milan, London, Paris, Spain, Stockholm), and South America (São Paulo). AWS plans to publish a roadmap for future availability. Users can choose hourly billing with no commitments or Reserved Instances for cost savings. AWS recommends using its Pricing Calculator with your specific workload patterns to estimate bills.

The combination of speed, cost efficiency, and an integrated data lake query engine mean Redshift RG instances could help the system cope with changing workloads from AI agents, which allow users to query data in natural language, AWS claims.

Andrew Warfield, AWS VP and Distinguished Engineer, said Redshift engineers had worked closely with the Graviton team for a few years, and the announcement was the first of an expected string of integrations.

He told The Register the performance improvement would help the data lakehouse system cope with the increasing demands of agent-driven workloads, as general business users query organizational data in natural language, rather than BI or data specialists using SQL.

"AI agents, because of the whole chain-of-reasoning structure, are often very interactive with the data, so they'll issue a query, they'll limit it, they'll look at initial results, they'll decide what to do next, and then they'll adjust and ramp up. We are seeing a really, really significant uptick in the query rate," Warfield said.

The issue was not like the "intern problem" when a user finds "there was a scheduled query that was SELECT * against a petabyte of data that was running once a week," he said. 

"The agents are actually much more capable of being frugal with the results that they ask of the database, but they're issuing a lot more queries, because they're able to iterate back and forth with thinking about a result and jumping to the next question."

There may be another reason AWS has been keen to bolster the performance of its flagship data warehouse/data lake system. Since it has a significant lead in the ever-expanding cloud market, AWS has seen customers store huge amounts of data in its S3 object storage. That gave Redshift a natural advantage for customers who wanted to work with data in the same environment. 

However, since AWS backed the Iceberg open table format in early 2023, it has become easier for customers to bring their analytics engine of choice to the data,  regardless of environment. AWS doubled down on Iceberg with the launch of the S3 Tables bucket type for storing data in the Apache Iceberg format and potentially making it more available to rival analytics engines.

"We have been very intentional with the work around Iceberg, and with S3 Tables," Warfield said. "Work on data migrations is incredibly painful. Foundational decisions on things like data representation and governance; those are decisions that you don't want to remake in the future, because they're things that involve a lot of change management and moving stuff around, and they really slow teams down. Our decision to anchor on Iceberg was squarely based on customers telling us that it was the format that they were moving to, and we wanted to make sure that we allowed customers to represent data in a way that allowed them to use whatever engine they want, whether it's one of ours or something else."

The field of analytics is expanding beyond the specialist players. For example, application giant SAP has long had data warehouse systems for its own data, but it recently bought Iceberg specialist Dremio in an effort to make SAP a home for AI agents querying data in environments from other vendors.

Warfield said: "We already have multiple of our own engines. We already work closely with open source communities. We want to make sure, at a storage and data level, that customers always have the freedom to bring the right tool to work with the data." ®