Kafka events connect the durable state owned by services with the asynchronous work owned by workers.
Topic contracts
workflows: build, reschedule, termination, deletion, and workflow terminal effects.jobs: dispatch and container execution work.job_logs: deterministic retained stdout and stderr events.analytics: workflow, job, and log aggregate inputs.
Event identity
Events carry stable identifiers or deterministic keys for replay-sensitive side effects. Workflow generation and job dispatch attempt fields provide additional stale-work protection.
Trace headers
OpenTelemetry propagation fields are injected into Kafka headers and extracted by consumers.
Compatibility
Event schema changes must account for records already present in Kafka. Additive changes are preferred, and consumers must not assume a producer and every consumer deploy simultaneously.
Delivery expectation
Events are delivered at least once. Consumers must remain idempotent, preconditioned, or deduplicated.