Потоковая обработка (Kafka, Flink)

Stream processing обрабатывает данные в реальном времени, критично для fraud detection, real-time analytics и personalization.

Ключевые системы

Apache Kafka: message broker, backbone streaming. Kafka Streams: stateful processing. Apache Flink: state-of-art, exactly-once, low-latency. Apache Beam: unified batch + streaming API. Storm, Spark Streaming (micro-batch). Choice зависит от use case complexity.

Semantic guarantees

At-most-once: fastest, могут теряться. At-least-once: возможны duplicates. Exactly-once: строгий, но complex. Flink native exactly-once. Kafka Streams: exactly-once с transactional producer. Critical для financial events (no duplicate bets, deposits).

State management

Stateful operators: aggregations, joins, windows. Backing store: RocksDB embedded. Checkpointing для fault tolerance. Savepoints для upgrades. Scale-out через parallelism. Memory management критичен для стабильности.

iGaming use cases

Real-time fraud scoring: event → risk score < 100ms. Session metrics: DAU, revenue в реальном времени. Behavioral triggers: RG interventions. Live odds update. Bonus qualification tracking. Push notifications personalization. All 24/7 operations.