Using a Modern Database with ML to Prevent Financial Fraud
summary
Zeta builds Himalaya feature store to prevent fraud at scale, serving low-latency lookups for 1M+ transactions and billions of events per day. It uses Flink/Kafka for streaming, Spark for batch, S3 for offline, and YugabyteDB as the online store; a 100M-record (1 KB) benchmark favors Yugabyte over CockroachDB. Features support windowed aggregates, composite index keys, protobuf-stored tumbling windows with TTL, online→offline sync, multi‑tenant cloud‑native deployment, and a planned Python SDK and gRPC support.
AI-written summary · about
chapters (13)
- 0:00Himalaya fraud prevention intro
- 1:08Agenda and fraud risk overview
- 2:10Real-time fraud system reqs
- 3:13Fraud risk management architecture
- 5:40Fraud detection challenges
- 8:09How Himalaya solves these challenges
- 10:37Himalaya user flow
- 13:05Window aggregate features
- 15:13Himalaya system architecture
- 16:38Himalaya tech stack
- 17:00Online data store reqs
- 19:48Database options comparison
- 21:13Himalaya future roadmap
related talks
18mBuilding a Scalable, Resilient System for Real-Time Fraud Detection
37mHow YugabyteDB Empowers AI-powered Fraud Intelligence
29mBuilding Resilient AI Platforms: Financial Fraud Detection
29mDelivering Always-On Financial Services with
22mBuilding a Modern Data Foundation: A Conversation with OCBC's Head of Data Architecture
32mDiscover the Modern Way to Build Real-Time Analytics Solutions at Any Scale
26mEfforts Toward Adoption and Uncovering Hidden Features
36mData Warehouse, Data Lakes, What's Next?