Speaker: Venkat Venkataramani

Speaker: Girish Baliga

Search Infrastructure powers search for all of Uber's products - Delivery, Mobility, Maps, Ads, U4B, Freight, and numerous internal products . We have built a custom search platform using Apache Lucene to address Uber's unique real-time business requirements. In this talk, we will present how we are able to support very high QPS traffic on a global scale with fresh data. We will also cover how we are innovating in the AI space with vector search using Apache Lucene’s KNN implementation.

Speaker: Shu Zhang

Pinterest's ad delivery system is a large-scale machine learning system that powers personalized ad recommendations for hundreds of millions of users with millisecond latency. In this talk, we will give an overview of the evolution of our ads ML infrastructure, and then dive deep into the real-time ML systems we have built. We will discuss the ranking systems that rank hundreds of millions of ad documents per second, as well as the toolkits we have developed to make these systems stable and performant. We will also share our insights on some of the key challenges we have faced, such as ML feature engineering, and discuss the trends we see in ML infrastructure at Pinterest and in the industry.

Speakers: Yexi Jiang, Michelle Gong

Recommendation systems are critical to Roblox. They power the homepage and other content discovery products, including ads, social, avatar, and more. In late 2022, we launched the real time recommendation for the homepage. The solution was then adopted by other S&D products at Roblox. In this talk, we start with delineating the problems and challenges of the homepage recommendation. Then, we delve into the solutions we implemented to support real-time recommendations for hundreds of millions of users. Lastly, we discuss the tools we developed to expedite the iteration process, reducing the iteration time from weeks to days.

Speaker: Jaya Kawale

Tubi is Fox’s video streaming platform and has millions of viewers per month. Personalized recommendations are a critical aspect of the user experience as they allow viewers to quickly find the content they would like to watch. In this talk, we will talk about the challenges of building a personalization service that moved from batch to real time, the data stack that we use to store and process embeddings, how we reduced the storage space of embeddings and how our datastore is extended to support hierarchical clustering of embeddings to get better personalization. We present measurements on how much we saved by moving to real time because it reduces the cost of unnecessary computation and storage.

Speaker: Emmanuel Fuentes

Speaker: Nikhil Garg

Quora is a social question-and-answer website with over four hundred million monthly active users.  In this talk, we describe how large scale real-time recommendation systems are built across the industry via the case study of building Quora’s personalized news feed. The talk will lay out the common architectural blueprint of modern recommendation systems, describe scaling challenges along performance, reliability, and cloud costs for each of the components in the blueprint. Finally, the talk will conclude with a brief discussion of the future trends in recommendation systems and how they are expected to influence the architecture of the next generation of recommendation systems.

Speaker: Kai Waehner