Problem
A recommender learns a member's intent from past actions, but the usual batch feature pipeline delays them by hours or days. Until the next batch run, the system keeps serving the old preference. In a simulation on LinkedIn's job recommender, a 1-hour delay lowers ROC-AUC by 3.51% against a 1-minute delay, and a 24-hour delay by 4.45%.
Example from the article: a member who applied for jobs in San Francisco a month ago applies for jobs in New York today. With a 24-hour delay, recommendations keep favoring San Francisco for another day.
Solution in brief
Keep the last 96 hours of member actions, with entity attributes already joined, in a low-latency online store. Compute features from them on demand at request time and combine them with batch features for the long term. Actions become available within 0.1 to 15 seconds, and retrieval takes under 50 ms.
Questions this case answers
- How to make a ranking model use user actions taken seconds ago when features normally come from a periodic batch pipeline
- How to serve features built from a short window of recent user events within a tight latency budget
- How to make recommendations react to a member's actions within seconds instead of hours or days?
- How to compute features from a user's recent actions on demand within 100 ms at serving time?
Tools and methods
- Apache Kafka — stream of action events
- Apache Samza (Samza SQL) — stream processor that filters events and joins attributes
- Apache Pinot — online store of recent actions with near real-time ingestion
- Venice — key-value store of entity attributes
- Apache Spark, HDFS — batch pipeline for long-term features and training data
Google or a one-time email code. No password.