In this talk, we will explore the challenges and strategies of tuning low latency online feature stores to tame the p99 latencies, shedding light on the importance of choosing the right data model. As modern machine learning applications require increasingly fast and efficient feature retrieval/computation, addressing p99 latencies has become a crucial aspect in maintaining system performance and user satisfaction.

We will begin by defining p99 latencies and their impact on online feature stores. We will then delve into common issues that contribute to high p99 latencies and their implications on system performance. By sharing insights into best practices for latency reduction, we will demonstrate the value of optimizing low latency online feature stores for better performance.

Central to our discussion will be the significance of selecting the appropriate data model to meet application demands. We will cover various data model alternatives, highlighting their respective strengths and weaknesses.

Our goal is to empower practitioners to make informed decisions and ultimately achieve a well-tuned, efficient feature store capable of taming p99 latencies.