The Challenges of Efficiently Scaling up Real-Time Bidding

Recently, Turn published two research papers at one of the most prestigious data mining conferences, the IEEE International Conference on Data Mining 2015 (ICDM). Lets discuss one of them —  a paper explaining how Turn handles the challenge of high volume traffic in real-time bidding.

Real-time bidding (RTB) refers to the buying and selling of online ads through programmatic, instantaneous auctions that occur in the order of milliseconds before a webpage is loaded by a consumer. As one of the largest demand-side platform (DSP) companies, currently handles over three million queries per second (QPS). At the time this paper was written, our QPS was 2.5 million.

In the paper, we present how we built a high-performance RTB platform inside Turn. The increased traffic to our platform has the similar effect of a Distributed Denial-of-Service attack (DDoS), i.e., at the peak time a large number of less valuable bid requests could potentially jam our platform and slow down our bidding.

We noticed that the QPS fluctuates dramatically within the day, and generally, only a small portion of bid requests have interested buyers. Given that we can satisfy most of our advertiser budgets with a smaller set of bid requests, we propose to process only the requests that are relevant. We discuss how to build a hierarchical selection model for bid requests, so the platform can intelligently throttle bid requests at the desired level without adversely impacting overall spending and performance.

Qualified ads are scored in a hierarchical way to ensure thorough estimation of potential outcomes within the given time latency. In the first tier of scoring, we select a set of promising ads for the future thorough scoring by analyzing a small set of features. Using limited information to select ads is a special case of Partially Observable Markov Decision Process (POMDP). Although the underlying dynamics of our problem are fully observable and Markovian, since our time constraint prohibits us from direct access to the complete information, our decision-making requires keeping track of (possibly) the entire history of the process, and aggregating the decisions/rewards into an efficient form. Instead of explicitly defining an exploration factor such as in ε-greedy or Boltzmann distribution to balance exploration and exploitation, we rely on the stochastically sampled rewards to do exploration. To enable efficient sampling, we model the value function as a continuous probability distribution. In the second tier of scoring, we then utilize logistic regression to combine predictions from multiple resources into the final one.

We empirically show that with this refactored architecture, our platform is able to win more auctions and bring more value to clients by stabilizing the bidding pipeline.

About Amobee

Founded in 2005, Amobee is an advertising platform that understands how people consume content. Our goal is to optimize outcomes for advertisers and media companies, while providing a better consumer experience. Through our platform, we help customers further their audience development, optimize their cross channel performance across all TV, connected TV, and digital media, and drive new customer growth through detailed analytics and reporting. Amobee is a wholly owned subsidiary of Tremor International, a collection of brands built to unite creativity, data and technology across the open internet.

If you’re curious to learn more, watch the on-demand demo or take a deep dive into our Research & Insights section where you can find recent webinars on-demand, media plan insights & activation templates, and more data-driven content. If you’re ready to take the next step into a sustainable, consumer-first advertising future, contact us today.