Evolution and Scaling of Feature Store at Uber
summary
Palette serves 21,000+ features at ~25M queries/sec with 99.999% uptime and tackles scaling by automating onboarding via JSON configs, mirroring Spark and Cassandra partitioning for efficient dispersal, and supporting Kafka, Spark, Hive, Cassandra, Flink, Redis and an in‑house Sia for embeddings. Key reliability gains come from batching and incremental materialization to avoid huge shuffles, an auto‑config module that tunes Spark resources, resumable jobs that persist HDFS checkpoints, delta/column and selective dispersal to cut writes, caching tiers, and time‑shift features for online/offline consistency.
AI-written summary · about
chapters (9)
- 0:00Introduction to feature store
- 5:07Architecture at scale
- 6:06Scalability challenges
- 6:39Automated feature onboarding
- 8:19Offline serving optimizations
- 16:15Online serving infrastructure
- 21:44Online and offline consistency
- 25:18Quality and metadata management
- 30:40Future directions and Q&A
related talks
33mEvolution and Scaling of Feature Store at Uber
25mAccelerating ML at Uber with the Palette Feature Store
30mHow Feature Stores Enable Operational ML
32mA feature Store for Machine Learning
08mFeature Store Usage Patterns - From a Single Data Scientist to an Enterprise
25mBuilding A Feature Store For Hyper Growth
33mFeathr, an Open Source, Battletested, Scalable and Enterprise-Grade Feature Store
23mHopsworks Feature Store after 4 years: Lessons learned and what's next