Heroku Reduces Operational Overhead by 90% Using Amazon DynamoDB

Founded in 2007, Heroku is one of the first cloud PaaS providers and supports a wide range of programming languages, including Ruby, Python, Node.js, PHP, .NET, and Java. Heroku’s PaaS offering is the chosen solution for millions of applications, from small startups to large enterprises. Heroku is built entirely on AWS. “A development team needs to have a set of tools that are inclusive of development and production with a standard way to build, test, deploy, and more,” says Bob Wise, CEO of Heroku. “We use many AWS services to provide these capabilities to developers, but we pull them together into a PaaS offering—a customer can use Heroku and access all the tools that they need.”

Previously, Heroku used self-managed Apache Cassandra clusters for its Metrics as a Service (MetaaS) system, which collects and aggregates application metrics. This system also sends alerts for performance issues and drives Heroku’s automatic scaling functionality. Over time, the company’s engineers were burdened by time-consuming management tasks associated with operating these clusters, which took their focus away from innovation and new feature development for the platform.

Heroku chose to migrate from self-managed Cassandra clusters to Amazon DynamoDB, a serverless, NoSQL database service. To streamline the migration process, the company also adopted AWS Database Migration Service (AWS DMS), a managed migration and replication service that helps move database and analytics workloads to AWS quickly, securely, and with minimal downtime and zero data loss.

“Our priorities were very much aligned with AWS: number one was security, and number two was operational stability,” says Wise. “We want to be as safe as possible, and we have a level of trust in the AWS team’s commitment to operating with security as the main principle.”

The original MetaaS system ingested hundreds of thousands of observations per second into Apache Kafka. These observations were then consumed by stream-processing jobs, which calculated various metrics for each application and published the time-series data back to Kafka. The final data was then written to a Cassandra database for long-term storage and querying.

After careful planning, Heroku created a parallel set of stream-processing jobs that wrote the same data from Kafka to Amazon DynamoDB without impacting the rest of the system. Heroku incrementally increased the percentage of queries reading from both Cassandra and Amazon DynamoDB until 100 percent were running through both code paths. Heroku implemented a series of tests to validate data consistency between the two databases and, after confirming that the data was in sync, the team began the final phase of the migration.

Starting with test environments, Heroku gradually cut over read queries from Cassandra to Amazon DynamoDB until they eventually read solely from Amazon DynamoDB. The migration proved successful, and by May 2023, 100 percent of queries to the MetaaS system were being served by Amazon DynamoDB. Because of this gradual approach, Heroku could complete the migration without impacting existing customers. “Unless you were closely watching page load times for our metrics tab, you didn’t notice a thing,” says Wise. “For a lot of the work we do here at Heroku, that’s the ultimate sign of success: no one even noticed.”