TechPrufer
Get started
Redis
Redis logoRedis · Feature trail

Real-time Fraud Detection

2 posts · Jun 2026 – Sep 2026

What this trail covers

This post details how real-time feature stores and sliding-window data structures fit into the scoring pipeline for fraud detection, covering scaling to billions of events and the importance of high availability. It emphasizes that fraud detection is primarily a latency problem, with tight constraints on scoring time. The post highlights the use of Redis for low-latency feature retrieval, employing sorted sets for exact velocity counts and probabilistic structures like Bloom filters, HyperLogLog, and Count-Min Sketch for memory-efficient counting. It also discusses the training-serving skew problem and the benefits of Redis as an in-memory feature store for online inference.

All Redis trails
  1. Delivering Real-Time Personalization with Databricks and Redis | RedisThis post details the integration of Databricks Real-Time Mode (RTM) with Redis for real-time personalization. RTM is introduced as a continuous processing engine for Apache Spark's Structured Streaming, enabling sub-second performance with strict p99 latencies. The post explains how RTM processes clickstream events, maintains per-user session state (recency-weighted scores), and scores candidate products. The output is then written to Redis using a ForeachWriter, storing recommendations as a sorted set (product to score) and a hash (product title and price), both with short TTLs. The website reads this data from Redis in a pipelined round trip per render for sub-millisecond serving. Performance benchmarks show under 160ms p99 from click to recommendation served at 100k events/s. The post also outlines the broader applicability of this compute-and-serve pattern across various use cases like fraud detection, agent coordination, and inventory management.
  2. Real-Time Fraud Detection: Latency, Features & ScaleThis post elaborates on the real-time fraud detection capability by detailing the critical role of latency constraints in the scoring pipeline, the challenges of feature retrieval for fraud models, and the training-serving skew problem. It specifically highlights the use of Redis sorted sets for exact velocity counts and probabilistic data structures (Bloom filters, HyperLogLog, Count-Min Sketch) for memory-efficient counting. The post also emphasizes the architectural separation of event ingestion, stream processing for state maintenance, and low-latency feature stores like Redis for online inference.