AWS Big Data Blog

Achieve peak performance and boost scalability using multiple Amazon Redshift serverless workgroups and Network Load Balancer

As data analytics use cases grow, factors of scalability and concurrency become crucial for businesses. Your analytic solution architecture should be able to handle large data volumes at high concurrency and without compromising speed, thereby delivering a scalable high-performance analytics environment.

Amazon Redshift Serverless provides a fully managed, petabyte-scale, auto scaling cloud data warehouse to support high-concurrency analytics. It offers data analysts, developers, and scientists a fast, flexible analytic environment to gain insights from their data with optimal price-performance. Redshift Serverless auto scales during usage spikes, enabling enterprises to cost-effectively help meet changing business demands. You can benefit from this simplicity without changing your existing analytics and business intelligence (BI) applications.

To help meet demanding performance needs like high concurrency, usage spikes, and fast query response times while optimizing costs, this post proposes using Redshift Serverless. The proposed solution aims to address three key performance requirements:

  • Support thousands of concurrent connections with high availability by using multiple Redshift Serverless endpoints behind a Network Load Balancer
  • Accommodate hundreds of concurrent queries with low-latency service level agreements through scalable and distributed workgroups
  • Enable subsecond response times for short queries against large datasets using the fast query processing of Amazon Redshift

The suggested architecture uses multiple Redshift Serverless endpoints accessed through a single Network Load Balancer client endpoint. The Network Load Balancer evenly distributes incoming requests across workgroups. This improves performance and reduces latency by scaling out resources to meet high throughput and low latency demands.

Solution overview

The following diagram outlines a Redshift Serverless architecture with multiple Amazon Redshift managed VPC endpoints behind a Network Load Balancer.

The following are the main components of this architecture:

  • Amazon Redshift data sharing – This allows you to securely share live data across Redshift clusters, workgroups, AWS accounts, and AWS Regions without manually moving or copying the data. Users can see up-to-date and consistent information in Amazon Redshift as soon as it’s updated. With Amazon Redshift data sharing, the ingestion can be done at the producer or consumer endpoint, allowing the other consumer endpoints to read and write the same data and thereby enabling horizontal scaling.
  • Network Load Balancer – This serves as the single point of contact for clients. The load balancer distributes incoming traffic across multiple targets, such as Redshift Serverless managed VPC endpoints. This increases the availability, scalability, and performance of your application. You can add one or more listeners to your load balancer. A listener checks for connection requests from clients, using the protocol and port that you configure, and forwards requests to a target group. A target group routes requests to one or more registered targets, such as Redshift Serverless managed VPC endpoints, using the protocol and the port number that you specify.
  • VPC – Redshift Serverless is provisioned in a VPC. By creating a Redshift managed VPC endpoint, you enable private access to Redshift Serverless from applications in another VPC. This design allows you to scale by having multiple VPCs as needed. The VPC endpoint provides a dedicate private IP for each Redshift Serverless workgroup to be used as the target groups on the Network Load Balancer.

Create an Amazon Redshift managed VPC endpoint

Complete the following steps to create the Amazon Redshift managed VPC endpoint:

  1. On the Redshift Serverless console, choose Workgroup configuration in the navigation pane.
  2. Choose a workgroup from the list.
  3. On the Data access tab, in the Redshift managed VPC endpoints section, choose Create endpoint.
  4. Enter the endpoint name. Create a name that is meaningful for your organization.
  5. The AWS account ID will be populated. This is your 12-digit account ID.
  6. Choose a VPC where the endpoint will be created.
  7. Choose a subnet ID. In the most common use case, this is a subnet where you have a client that you want to connect to your Redshift Serverless instance.
  8. Choose which VPC security groups to add. Each security group acts as a virtual firewall to control inbound and outbound traffic to resources protected by the security group, such as specific virtual desktop instances.

The following screenshot shows an example of this workgroup. Note down the IP address to use during the creation of the target group.

Repeat these steps to create all your Redshift Serverless workgroups.

Add VPC endpoints for the target group for the Network Load Balancer

To add these VPC endpoints to the target group for the Network Load Balancer using Amazon Elastic Compute Cloud (Amazon EC2), complete the following steps:

  1. On the Amazon EC2 console, choose Target groups under Load Balancing in the navigation pane.
  2. Choose Create target group.
  3. For Choose a target type, select Instances to register targets by instance ID, or select IP addresses to register targets by IP address.
  4. For Target group name, enter a name for the target group.
  5. For Protocol, choose TCP or TCP_UDP.
  6. For Port, use 5439 (Amazon Redshift port).
  7. For IP address type, choose IPv4 or IPv6. This option is available only if the target type is Instances or IP addresses and the protocol is TCP or TLS.
  8. You must associate an IPv6 target group with a dual-stack load balancer. All targets in the target group must have the same IP address type. You can’t change the IP address type of a target group after you create it.
  9. For VPC, choose the VPC with the targets to register.
  10. Leave the default selections for the Health checks section, Attributes section, and Tags section.

Create a load balancer

After you create the target group, you can create your load balancer. We recommend using port 5439 (Amazon Redshift default port) for it.

The Network Load Balancer serves as a single-access endpoint and will be used on connections to reach Amazon Redshift. This allows you to add more Redshift Serverless workgroups and increase the concurrency transparently.

Testing the solution

We tested this architecture to run three BI reports with the TPC-DS dataset (cloud benchmark dataset) as our data. Amazon Redshift includes this dataset for free when you choose to load sample data (sample_data_dev database). The installation also provides the queries to test the setup.

Among all the queries from TPC-DS benchmark, we chose the following three to use as our report queries. We changed the first two report queries to use a CREATE TABLE AS SELECT (CTAS) query on temporary tables instead of the WITH clause to emulate options you can see on a typical BI tool. For our testing, we also disabled the result cache to make sure that Amazon Redshift would run the queries every time.

The set of queries contains the creation of temporary tables, a join between those tables, and the cleanup. The cleanup step drops tables. This isn’t needed because they’re deleted at the end of the session, but this aims to simulate all that the BI tool does.

We used Apache JMETER to simulate clients invoking the requests. To learn more about how to use and configure Apache JMETER with Amazon Redshift, refer to Building high-quality benchmark tests for Amazon Redshift using Apache JMeter.

For the tests, we used the following configurations:

  • Test 1 – A single 96 RPU Redshift Serverless vs. three workgroups at 32 RPU each
  • Test 2 – A single 48 RPU Redshift Serverless vs. three workgroups at 16 RPU each

We tested three reports by spawning 100 sessions per report (300 total). There were 14 statements across the three reports (4,200 total). All sessions were triggered simultaneously.

The following table summarizes the tables used in the test.

Table Name Row Count
Catalog_page 93,744
Catalog_sales 23,064,768
Customer_address 50,000
Customer 100,000
Date_dim 73,049
Item 144,000
Promotion 2,400
Store_returns 4,600,224
Store_sales 46,086,464
Store 96
Web_returns 1,148,208
Web_sales 11,510,144
Web_site 240

Some tables were modified by ingesting more data than what the TPC-DS schema offers on Amazon Redshift. Data was reinserted on the table to increase the size.

Test results

The following table summarizes our test results.

TEST 1 . Time Consumed Number of Queries Cost Max Scaled RPU Performance
Single: 96 RPUs 0:02:06 2,100 $6 279 Base
Parallel: 3x 32 RPUs 0:01:06 2,100 $1.20 96 48.03%
Parallel 1 (32 RPU) 0:01:03 688 $0.40 32 50.10%
Parallel 2 (32 RPU) 0:01:03 703 $0.40 32 50.13%
Parallel 3 (32 RPU) 0:01:06 709 $0.40 32 48.03%
TEST 2 . Time Consumed Number of Queries Cost Max Scaled RPU Performance
Single: 48 RPUs 0:01:55 2,100 $3.30 168 Base
Parallel: 3x 16 RPUs 0:01:47 2,100 $1.90 96 6.77%
Parallel 1 (16 RPU) 0:01:47 712 $0.70 36 6.77%
Parallel 2 (16 RPU) 0:01:44 696 $0.50 25 9.13%
Parallel 3 (16 RPU) 0:01:46 692 $0.70 35 7.79%

The preceding table shows that the parallel setup was faster than the single at a lower cost. Also, in our tests, even though Test 1 had double the capacity of Test 2 for the parallel setup, the cost was still 36% lower and the speed was 39% faster. Based on these results, we can conclude that for workloads that have high throughput (I/O), low latency, and high concurrency requirements, this architecture is cost-efficient and performant. Refer to the AWS Pricing Cost Calculator for Network Load Balancer and VPC endpoints pricing.

Redshift Serverless automatically scales the capacity to deliver optimal performance during periods of peak workloads including spikes in concurrency of the workload. This is evident from the maximum scaled RPU results in the preceding table.

Recently released features of Redshift Serverless such as MaxRPU and AI-driven scaling were not used for this test. These new features can increase the price-performance of the workload even further.

We recommend enabling cross-zone load balancing on the Network Load Balancer because it distributes requests from clients to registered targets. Enabling cross-zone load balancing will help balance the requests among the Redshift Serverless managed VPC endpoints irrespective of the Availability Zone they are configured in. Also, if the Network Load Balancer receives traffic from only one server (same IP), you should always use an odd number of Redshift Serverless managed VPC endpoints behind the Network Load Balancer.


In this post, we discussed a scalable architecture that increases the throughput of Redshift Serverless in low latency, high concurrency scenarios. Having multiple Redshift Serverless workgroups behind a Network Load Balancer can deliver a horizontally scalable solution at the best price-performance.

Additionally, Redshift Serverless uses AI techniques (currently in preview) to scale automatically with workload changes across all key dimensions—such as data volume changes, concurrent users, and query complexity—to meet and maintain your price-performance targets.

We hope this post provides you with valuable guidance. We welcome any thoughts or questions in the comments section.

About the Authors

Ricardo Serafim is a Senior Analytics Specialist Solutions Architect at AWS.

Harshida Patel is a Analytics Specialist Principal Solutions Architect, with AWS.

Urvish Shah is a Senior Database Engineer at Amazon Redshift. He has more than a decade of experience working on databases, data warehousing and in analytics space. Outside of work, he enjoys cooking, travelling and spending time with his daughter.

Amol Gaikaiwari is a Sr. Redshift Specialist focused on helping customers realize their business outcomes with optimal Redshift price-performance. He loves to simplify data pipelines and enhance capabilities through adoption of latest Redshift features.