AWS Big Data Blog

How SafetyCulture scales unpredictable dbt Cloud workloads in a cost-effective manner with Amazon Redshift

This post is co-written by Anish Moorjani, Data Engineer at SafetyCulture.

SafetyCulture is a global technology company that puts the power of continuous improvement into everyone’s hands. Its operations platform unlocks the power of observation at scale, giving leaders visibility and workers a voice in driving quality, efficiency, and safety improvements.

Amazon Redshift is a fully managed data warehouse service that tens of thousands of customers use to manage analytics at scale. Together with price-performance, Amazon Redshift enables you to use your data to acquire new insights for your business and customers while keeping costs low.

In this post, we share the solution SafetyCulture used to scale unpredictable dbt Cloud workloads in a cost-effective manner with Amazon Redshift.

Use case

SafetyCulture runs an Amazon Redshift provisioned cluster to support unpredictable and predictable workloads. A source of unpredictable workloads is dbt Cloud, which SafetyCulture uses to manage data transformations in the form of models. Whenever models are created or modified, a dbt Cloud CI job is triggered to test the models by materializing the models in Amazon Redshift. To balance the needs of unpredictable and predictable workloads, SafetyCulture used Amazon Redshift workload management (WLM) to flexibly manage workload priorities.

With plans for further growth in dbt Cloud workloads, SafetyCulture needed a solution that does the following:

  • Caters for unpredictable workloads in a cost-effective manner
  • Separates unpredictable workloads from predictable workloads to scale compute resources independently
  • Continues to allow models to be created and modified based on production data

Solution overview

The solution SafetyCulture used is comprised of Amazon Redshift Serverless and Amazon Redshift Data Sharing, along with the existing Amazon Redshift provisioned cluster.

Amazon Redshift Serverless caters to unpredictable workloads in a cost-effective manner because compute cost is not incurred when there is no workload. You pay only for what you use. In addition, moving unpredictable workloads into a separate Amazon Redshift data warehouse allows each Amazon Redshift data warehouse to scale resources independently.

Amazon Redshift Data Sharing enables data access across Amazon Redshift data warehouses without having to copy or move data. Therefore, when a workload is moved from one Amazon Redshift data warehouse to another, the workload can continue to access data in the initial Amazon Redshift data warehouse.

The following figure shows the solution and workflow steps:

  1. We create a serverless instance to cater for unpredictable workloads. Refer to Managing Amazon Redshift Serverless using the console for setup steps.
  2. We create a datashare called prod_datashare to allow the serverless instance access to data in the provisioned cluster. Refer to Getting started data sharing using the console for setup steps. Database names are identical to allow queries with full path notation database_name.schema_name.object_name to run seamlessly in both data warehouses.
  3. dbt Cloud connects to the serverless instance and models, created or modified, are tested by being materialized in the default database dev, in either each users’ personal schema or a pull request related schema. Instead of dev, you can use a different database designated for testing. Refer to Connect dbt Cloud to Redshift for setup steps.
  4. You can query materialized models in the serverless instance with materialized models in the provisioned cluster to validate changes. After you validate the changes, you can implement models in the serverless instance in the provisioned cluster.

Outcome

SafetyCulture carried out the steps to create the serverless instance and datashare, with integration to dbt Cloud, with ease. SafetyCulture also successfully ran its dbt project with all seeds, models, and snapshots materialized into the serverless instance via run commands from the dbt Cloud IDE and dbt Cloud CI jobs.

Regarding performance, SafetyCulture observed dbt Cloud workloads completing on average 60% faster in the serverless instance. Better performance could be attributed to two areas:

  • Amazon Redshift Serverless measures compute capacity using Redshift Processing Units (RPUs). Because it costs the same to run 64 RPUs in 10 minutes and 128 RPUs in 5 minutes, having a higher number of RPUs to complete a workload sooner was preferred.
  • With dbt Cloud workloads isolated on the serverless instance, dbt Cloud was configured with more threads to allow materialization of more models at once.

To determine cost, you can perform an estimation. 128 RPUs provides approximately the same amount of memory that an ra3.4xlarge 21-node provisioned cluster provides. In US East (N. Virginia), the cost of running a serverless instance with 128 RPUs is $48 hourly ($0.375 per RPU hour * 128 RPUs). In the same Region, the cost of running an ra3.4xlarge 21-node provisioned cluster on demand is $68.46 hourly ($3.26 per node hour * 21 nodes). Therefore, an accumulated hour of unpredictable workloads on a serverless instance is 29% more cost-effective than an on-demand provisioned cluster. Calculations in this example should be recalculated when performing future cost estimations because prices may change over time.

Learnings

SafetyCulture had two key learnings to better integrate dbt with Amazon Redshift, which can be helpful for similar implementations.

First, when integrating dbt with an Amazon Redshift datashare, configure INCLUDENEW=True to ease management of database objects in a schema:

ALTER DATASHARE datashare_name SET INCLUDENEW = TRUE FOR SCHEMA schema;

For example, assume the model customers.sql is materialized by dbt as the view customers. Next, customers is added to a datashare. When customers.sql is modified and rematerialized by dbt, dbt creates a new view with a temporary name, drops customers, and renames the new view to customers. Although the new view carries the same name, it’s a new database object that wasn’t added to the datashare. Therefore, customers is no longer found in the datashare.

Configuring INCLUDENEW=True allows new database objects to be automatically added to the datashare. An alternative to configuring INCLUDENEW=True and providing more granular control is the use of dbt post-hook.

Second, when integrating dbt with more than one Amazon Redshift data warehouse, define sources with database to aid dbt in evaluating the right database.

For example, assume a dbt project is used across two dbt Cloud environments to isolate production and test workloads. The dbt Cloud environment for production workloads is configured with the default database prod_db and connects to a provisioned cluster. The dbt Cloud environment for test workloads is configured with the default database dev and connects to a serverless instance. In addition, the provisioned cluster contains the table prod_db.raw_data.sales, which is made available to the serverless instance via a datashare as prod_db′.raw_data.sales.

When dbt compiles a model containing the source {{ source('raw_data', 'sales') }}, the source is evaluated as database.raw_data.sales. If database is not defined for sources, dbt sets the database to the configured environment’s default database. Therefore, the dbt Cloud environment connecting to the provisioned cluster evaluates the source as prod_db.raw_data.sales, while the dbt Cloud environment connecting to the serverless instance evaluates the source as dev.raw_data.sales, which is incorrect.

Defining database for sources allows dbt to consistently evaluate the right database across different dbt Cloud environments, because it removes ambiguity.

Conclusion

After testing Amazon Redshift Serverless and Data Sharing, SafetyCulture is satisfied with the result and has started productionalizing the solution.

“The PoC showed the vast potential of Redshift Serverless in our infrastructure,” says Thiago Baldim, Data Engineer Team Lead at SafetyCulture. “We could migrate our pipelines to support Redshift Serverless with simple changes to the standards we were using in our dbt. The outcome provided a clear picture of the potential implementations we could do, decoupling the workload entirely by teams and users and providing the right level of computation power that is fast and reliable.”

Although this post specifically targets unpredictable workloads from dbt Cloud, the solution is also relevant for other unpredictable workloads, including ad hoc queries from dashboards. Start exploring Amazon Redshift Serverless for your unpredictable workloads today.


About the authors

Anish Moorjani is a Data Engineer in the Data and Analytics team at SafetyCulture. He helps SafetyCulture’s analytics infrastructure scale with the exponential increase in the volume and variety of data.

Randy Chng is an Analytics Solutions Architect at Amazon Web Services. He works with customers to accelerate the solution of their key business problems.