Exactly-Once Semantics Auditor

Audit the following streaming pipeline's claim of "exactly once." Treat framework marketing claims as unproven until the source, processor, side effects, and sink are analyzed together. Architecture: [source → broker → processor → sinks] Event identity and keys: [details] Checkpoint and state backend: [configuration] Sink write behavior: [transactions / upserts / append] External side effects: [APIs, notifications, secondary stores] Failure and replay behavior observed: [details] Business correctness requirement: [what duplication or loss means] Return: 1. A delivery-semantics matrix for every boundary 2. Counterexamples showing where duplicates or loss can still occur 3. Assessment of source replayability and deterministic processing 4. Transaction, checkpoint, and sink-commit coordination analysis 5. Strategy for idempotent external side effects 6. Stable deduplication key and state-retention recommendation 7. Failure-injection tests for crashes before/after each commit boundary 8. Honest final classification: at-most-once, at-least-once, effectively-once, or end-to-end exactly-once 9. Minimal changes required to meet the stated business guarantee Use a timeline for at least two difficult crash scenarios. State which guarantees fail after the deduplication retention window expires.