Handling Five Billion Ad Requests Daily at Low Latency Using Amazon EMR with Siprocal

In September 2021, Siprocal’s relatively small 10-person engineering team began the migration alongside TO THE NEW, which has been an AWS Partner since 2013. The team created a parallel setup on AWS and replicated the components of the Apache Hadoop cluster, which consisted of 40 nodes running 32 processors each. Engineers slowly migrated 1 PB of data that had been stored on disk within the Apache Hadoop cluster to Amazon Simple Storage Service (Amazon S3), an object storage service that offers industry-leading scalability, data availability, security, and performance. Siprocal completed the migration with no downtime or loss of data. “We went from a very old infrastructure to AWS in 8 months,” says Wu. “This is a big achievement.”

Siprocal used the migration as an opportunity to improve the efficiency of its data pipeline. It now directs incoming ad requests using Application Load Balancer, which provides advanced request routing targeted at delivery of modern application architectures. For example, Siprocal can route traffic to a user interface, where clients and internal teams manage ad campaigns. Additionally, an event service receives impressions from publishers and connects them to various databases for cached data, metadata, an event message bus, and the ad-pacing service that manages the rate at which ad campaigns roll out.

Business-critical applications run in Amazon Elastic Container Service (Amazon ECS), a fully managed container orchestration service that simplifies deployment, management, and scaling of containerized applications. Containers run using AWS Fargate, a serverless, pay-as-you-go compute engine that lets organizations focus on building applications without managing servers. “Previously, we didn’t take in more business because we just couldn’t support it,” says Steven Cothren, director of software engineering at Siprocal. “Now we can scale to meet the business demand.” Before the migration Siprocal could handle 1.5 billion ad requests per day, but now it can process up to 5 billion ad requests per day during peak times, and the average response time is 300 ms. “The speed at which we respond to requests directly affects our revenue,” Cothren says. “A component of determining which ad gets displayed is who responds quickest. Our performance using AWS helps with that. It also avoids outages. If we have an outage, we make no money.”

Freed from storage limitations for logs, Siprocal improves observability using Amazon CloudWatch, which collects and visualizes near-real-time logs, metrics, and event data in automated dashboards to streamline infrastructure and application maintenance. Engineers see the number of requests, application errors, response times, and other metrics. “Even if there’s not a problem yet, we can see if we’re trending toward a problem,” says Cothren. “We have a true view into all our applications.”