Amazon Redshift

What it is

Amazon Redshift is AWS’s cloud data warehouse, sold in two shapes. Provisioned clusters run on RA3 nodes you size and pay for by the hour, with storage billed separately as Redshift Managed Storage. Redshift Serverless bills compute in Redshift Processing Units (RPUs) only while queries run, with a base capacity you set between 4 and 512 RPUs (1,024 in some regions). Both query the warehouse and, through Spectrum, data sitting in S3. Since 31 August 2026 Redshift also reads and writes Apache Iceberg v3 tables, which matters if your data lake is the long-term store and the warehouse is one engine among several.

The 2026 change that matters to ops teams is agent access. On 27 August 2026 AWS integrated Redshift with Agent Toolkit for AWS: the AWS MCP Server plus a set of Redshift skills (SQL syntax, metadata discovery, data loading, materialized views, migration playbooks) packaged as the aws-data-analytics plugin for Claude Code, Kiro and Cursor. It works against provisioned clusters and Serverless workgroups and carries no added charge. Separately, AWS publishes an open-source Redshift MCP server that runs SQL in read-only mode with validation checks against harmful statements. Inside the console, Amazon Q generative SQL in Query Editor v2 has been GA since September 2024 and turns plain-English questions into SQL once an admin enables it.

Why ops teams end up here

Almost nobody picks Redshift for RevOps. They inherit it: engineering standardised on AWS years ago and the product data already lives there. The real question is we’re on Redshift, not Snowflake. Do the warehouse-native GTM and comp tools still work, and can agents query it safely?

Mostly yes. Hightouch syncs from Redshift, Fivetran loads into it, Sigma queries it live, and both Everstage and Fullcast list it as a native connector. Zero-ETL integrations, GA since December 2024, replicate Salesforce and other SaaS applications into Redshift without a pipeline tool. They are billed through AWS Glue, so a team with only a few sources may not need Fivetran at all. A concrete use: a CS team asks Claude Code, through the MCP server, which renewals over $50,000 show falling weekly active users, and gets an answer straight from product telemetry and Salesforce data in the same warehouse.

Pricing reality

There is no seat price. List prices in US East (N. Virginia):

• Serverless: $0.375 per RPU-hour, billed per second with a 60-second minimum, so the 4-RPU floor costs $1.50 an hour while active. New Serverless users get $300 in credits that expire after 90 days. Serverless Reservations cut the rate by up to 24% on a one-year term and up to 45% on three years.
• Provisioned: ra3.xlplus at $1.086 per node-hour and ra3.4xlarge at $3.26, plus managed storage at $0.024 per GB-month. Each cluster earns one free hour of Concurrency Scaling per day, banking up to 30 hours.
• Spectrum: $5 per TB scanned in S3.

A worked band: a RevOps warehouse on Serverless with an 8-RPU base, active about four hours each business day, holding 500 GB. Compute comes to about $264 a month ($3 an hour × 4 hours × 22 days), storage to about $12. Roughly $280 a month, about $3,300 a year. The same workload on two ra3.xlplus nodes left running around the clock costs about $1,590 a month on demand, which is why small ops workloads belong on Serverless.

Best for

A RevOps or CS analytics lead at a company whose product data already lives in AWS, who wants warehouse-native activation, comp and BI tools on top of it without a migration project. It also suits teams that must keep customer data inside an AWS account and region for security review, and that want coding agents to query the warehouse under IAM rather than through a third-party service.

Not for

A company choosing its first warehouse with no AWS commitment. Snowflake and BigQuery have the larger GTM ecosystems and their own conversational analytics layers. Redshift’s AI story is mostly agent access through MCP, not a business-user chat product. If the main question is “do we need a warehouse or a CDP at all,” start with data warehouse vs CDP.

Versus the alternatives

• Snowflake — the share leader for GTM and the default target for every reverse-ETL and comp vendor. Pick it when you have no cloud lock-in and want CoWork-style chat for business users on top of the warehouse.
• Google BigQuery — the other share leader. Pick it when the company runs on Google Cloud or Workspace and wants per-TiB on-demand pricing with no clusters at all.
• Databricks — the fastest-growing entrant: press coverage in August 2026 put its data-warehousing run rate above $1.5 billion, up more than 100% year over year. Pick it when data science and ML already run on the lakehouse and you want Genie on the same data.
• Stay on Redshift when the data is already there. Migrating a working warehouse to get a slightly better GTM connector is rarely worth a quarter of engineering time.

Watch-outs

• Serverless defaults to a large base capacity. The console proposes 128 RPUs, which is $48 an hour while queries run. Guard: set the base to 8 RPUs for ops workloads, and add a usage limit on RPU-hours per day with an alert before it hits.
• BI tools and syncs keep Serverless awake. A dashboard refreshing every five minutes or an hourly reverse-ETL sync bills compute all day. Guard: schedule syncs into a few batch windows, materialize the models Sigma reads, and review RPU usage by workgroup weekly for the first month.
• Agent write access is one IAM policy away. The AWS MCP Server runs authenticated AWS API calls as the user it is given. Guard: give agents a dedicated IAM role and database user with SELECT-only grants on the modelled schemas, and use the open-source Redshift MCP server’s read-only mode for ad-hoc questions.
• Python UDFs are gone. Support ended after 30 June 2026, and older dbt packages and tutorials still use them. Guard: search your models for LANGUAGE plpythonu and port those functions to Lambda UDFs before relying on any package that ships them.
• Writeback tools want bigger nodes. Sigma recommends ra3.4xlarge or larger for its input tables on Redshift. Guard: test writeback on Serverless or your actual node size during the trial before you build an app around it.

Related: the warehouse-native GTM stack shows where a warehouse like Redshift sits next to Hightouch, Clay and the CRM.