Tableau Uses AWS to Scale Faster and Increase Resiliency of SaaS Offering

Tableau didn’t waste any time with its lift-and-shift Tableau Online migration. “We did one clean swoop to move everything—all the customer workloads and processes for Tableau Online—to AWS,” says Pankaj Dhingra, vice president of cloud engineering at Tableau. First, the migration team designed its network layer, providing security and connectivity among its environments on Amazon Virtual Private Cloud (Amazon VPC), which helps businesses launch AWS resources in logically isolated virtual networks that they define.

The next step of the migration involved Tableau Online’s storage layer. The company opted for Amazon Relational Database Service (Amazon RDS) for PostgreSQL, which makes it simple to set up, operate, and scale PostgreSQL deployments in the cloud. For compute, Tableau Online relied on Amazon Elastic Compute Cloud (Amazon EC2), a web service that provides secure, resizable compute capacity in the cloud. After discovering that the cloud-native file system on Windows File Server didn’t meet its performance needs, Tableau decided to use a distributed replicated storage solution backed by performance-enhancing Amazon Elastic Block Storage (Amazon EBS) volumes, which are durable block-level storage devices that can be attached to one instance or multiple instances at the same time. Amazon EBS uses Amazon EC2 instances to handle throughput and transaction-intensive workloads at any scale. As a result, queries into Tableau’s database, Hyper, became four times faster.

“Migrating to AWS was seamless,” says Dhingra, “and customers started having three to five times better performance across the board.” The longest load time (99th percentile) for visualizations dropped from 6.6 to 2.2 seconds, and shorter load times (50th percentile) dropped from 0.8 to 0.18 seconds. In one example, the median viz load time decreased from 41.10 seconds to 18.62 seconds, and the median load time distribution in the US-Seattle region decreased from 65.9 seconds to 42.6 seconds. “We got excellent help and resources from the AWS team,” says Dhingra. “AWS Solutions Architects helped us review our designs and future-proof the things that we were doing.” To make sure the cloud solution would perform as efficiently as possible, Tableau tested more than 20,000 Tableau vizzes, data import speeds, compute power, extract refreshes, and data center access speeds. Ultimately, the migration enabled Tableau to triple in scale: for example, the PoD in the US East (N. Virginia) region can support about 100,000 users, whereas the old architecture could handle 30,000–40,000 users per PoD.

The company also gained redundancy and georeplication at high-bandwidth, low-latency networking between AWS Regions and Availability Zones. Visualization load times for Tableau Online improved by two to three times, an observation that finalized Tableau’s decision to go all in on AWS. The company can bring up a new PoD in a new geography in a matter of days using AWS, compared to months on premises. In addition to launching the US East region PoD for Tableau Online in 2019, the company launched two PoDs in 2020: one in the Asia Pacific (Southeast-2) region and one in the Asia Pacific (Northeast-1) region.

In early 2017, Tableau’s foundational systems—which included a grid consisting of thousands of virtual machines for developers to develop and test code, as well as performance testing labs for benchmarking and performance testing to ensure high-quality and trustworthy software—still ran on on-premises data centers. This required millions of dollars in new hardware every quarter. And the company was budgeting for peak load capacity, even though almost 80 percent of the compute capacity remained unused at night. To gain flexibility, Tableau had purchased hundreds of racks, slotted thousands of servers, and kept building data center skills to manage it all. But the company wanted a fully managed, cost-effective solution—one that would not only provide flexibility and agility but also scale compute capacity quickly.

“Our developer productivity system needs were growing at a fast clip, and we spent most of our time procuring new hardware and adding it to our data centers,” says Dhingra. “And by the time we made it ready and functional, we had already outgrown that capacity.”

Tableau realized that although AWS would cost the same as its data centers, it would be more beneficial because AWS managed services would save Tableau time (including a 20 percent time reduction by the operations team) and resources by maintaining hardware, including patching and updates.

By moving to AWS, Tableau adopted a standardized compute solution that can be upgraded to the newest versions while enabling quick experiments to determine the exact compute the company needs. Tableau used this experimenting capability to rightsize Amazon EC2 instances for different workloads. For example, within Tableau Online, rightsizing saved the company about $1 million in 2019. Now, Tableau can do the equivalent of adding hardware in an on-premises system—but on AWS, the company can add provisioning much more quickly and without interrupting service to internal or external customers. In 2020, Tableau Online consistently exceeded its 99.9 percent availability target.

On AWS, Tableau increased the performance, reliability, and scalability of Tableau Online while simultaneously reducing costs and streamlining operations. The AWS-powered Tableau Online compute environment also challenged customers’ perceptions of cloud versus on-premises capabilities. “When customers tried Tableau Online, they found the performance to be better than their on-premises hosted Tableau Server,” says Dhingra. “Even though Tableau Online is a multitenant environment, we were able to guarantee better performance.”