Model calls are slow, expensive, rate-limited and non-deterministic compared with the rest of a stream-processing pipeline. Putting them into event-driven systems naively produces duplicate charges, poison messages, consumer-lag spirals and results nobody can reproduce. Done well, events are the most natural way to run AI at scale: enrichment, classification, embedding refresh, document processing and agent workflows all become asynchronous, replayable and observable.
This course starts from messaging fundamentals as they apply to AI: delivery semantics, partitioning and ordering, consumer groups versus Kafka share groups (Queues for Kafka, production-ready in Kafka 4.2), and when a classic queue such as RabbitMQ, SQS, Pub/Sub, Service Bus or NATS JetStream fits better. You then build AI consumers with idempotency keys, retries, dead-letter topics, backpressure and provider rate-limit handling, and use batch APIs where latency allows. You run model inference inside stream processors with Flink SQL ML_PREDICT and managed Flink AI functions, keep vector indexes fresh with Debezium CDC, and design schemas and contracts with Schema Registry, CloudEvents and AsyncAPI. The final modules cover agentic and long-running work (durable workflows with Temporal, event-driven agents, human-in-the-loop steps), replay and reprocessing when models change, observability and cost control, and data protection in streams. Product capabilities are stated as of October 2026.