Redis fraud feature store: MongoDB offline store on the left, Redis online store on the right, an RDI stream-sync belt between them, and a center verdict that flips green or red with the sub-millisecond Redis lookup time.
Fraud Feature Store
real-time inference on the Redis online store
connecting…
MongoDB
offline store
RDI
CDC stream-sync · lag —ms
Inference
decision · O(1)
offline feature store
batch · model training
2,000 card_features
500 merchant_features
300 terminal_features
The model trains here. Nothing serves inference from here, RDI mirrors the latest state into Redis.
SCORING NOW
$0
—
· · ·
Redis lookup— msO(1) · last-write-wins · no TTL
online store · O(1)
last-write-wins · no TTL · in-memory
throughput
0 txn/s
p99 lookup
0.0 ms
fraud blocked
0
Live RDI propagationchange the offline store, watch it reach Redisdemo card card_000001 · the app writes to MongoDB only, never to Redis
1 · MongoDB offline
idle
2 · RDI sync (CDC)
idle
3 · Redis online · decision
idle
real-time scoring · O(1) reads from the Redis online store · ~12% fraud in the stream