Reliable Messaging
Delivery. Follow the affected business state, message identity, publish, broker acceptance, consume, durable effect and acknowledgement through recovery. State every guarantee with its start, commit, end and accepted loss/duplicate window. Broker deduplication is not an end-to-end exactly-once business effect.
Choose depth and authority
Review/diagnosis reconstructs the affected boundary read-only and reports the supported cause, minimal correction and falsifier. Design defines the in-scope contracts at the requested fidelity. Build/fix changes the local owner and runs applicable checks, repairing introduced failures. Production publication, broker configuration, replay/redrive, purge/deletion, credentials and load/fault tests require exact environment/action/range/bounds authorization, preflight and fresh readback. A proposed recovery plan is not permission to execute it.
Preserve accepted decisions outside the requested change and explicit artifact or test limits. A narrow acknowledgement fix does not trigger a full redesign. Full readiness, migration or replay claims do require the complete relevant lifecycle. Additional safety-critical dependencies must be named; unrelated concerns do not become completion gates. Ask only for a material missing decision or authority.
Load the selected broker contract
Read the reference for the actual broker and relevant concern, verifying material version-specific behavior against current primary documentation:
- Kafka: acceptance, offsets, transactions, partition order, retention and replay.
- RabbitMQ: routing/confirms, acknowledgements, prefetch, quorum/poison handling and recovery.
- SQS: Standard/FIFO, identity, visibility/deletion and redrive.
- JetStream: publish ACK, consumer progress, retention, redelivery and replay.
- Operations: selected ordering, backpressure, quarantine, recovery, migration, security or capacity questions.
An unfamiliar broker needs its own documented contract, not a guess from one of these examples. Reference checks are conditional on the claimed guarantee; they do not authorize live operations or require reading all broker documents.
Frame identity and durable intent
Name the protected effect, relevant loss/duplicate tolerance, latency, ordering, fan-out and recovery horizon. When the architecture decision is open, compare synchronous calls or a single-store job with the broker's needed buffering, isolation, fan-out or replay. Do not repeat that comparison for a settled broker's narrow bug fix.
Classify event, command or competing-worker item. Use one ledger row per interacting publish/consume boundary, or just the affected row for a narrow task:
| Boundary | Guarantee/start/end | Identity/scope | Durable commit | Ambiguity | Recovery owner | Signal | Falsifier |
|---|
A logical message ID survives publish retry, redelivery, redrive and replay; offsets, delivery tags, receipt handles and attempt IDs are not substitutes. Retain envelope fields that affect behavior: type/version, tenant, ordering/causation, immutable intent and safe trace references. Scope idempotency by stable logical effect and tenant, not mutable deployment or consumer instance. A correlation ID is not a deduplication key. A new dedup scheme must account for already-applied effects before replay, or it can reapply the very effect it should preserve.
Publish and commit the effect
When business state and publish intent share a database, commit state and outbox together. Publish with the selected durable-acceptance contract, then mark relay completion; lost responses remain ambiguous under the same identity. If no atomic intent boundary exists, name the loss/duplicate window and reconciliation owner. Bound publisher queues and in-flight work. Broker transactions do not join an unrelated business database transaction.
For a transactional consumer, commit business effect and uniquely enforced inbox/effect identity together. For an external effect, propagate the stable idempotency key or durable downstream intent. Acknowledge/delete/commit recovery progress only after the relevant durable effect. Concurrent completions cannot acknowledge past unfinished work in an ordered lane. A preflight dedup read alone leaves a race.
Effect-before-ACK crashes and lost ACK responses must converge on durable state without multiplying effects. Retain dedup information across the longest redelivery, redrive, retention and offline-replay horizon; never-repeatable effects may require a permanent business key. Keep schema compatibility and retirement conditions for the producers/consumers that can coexist.
Prove the requested boundary
Use proof receipt as an optional coverage aid when several artifacts or guarantees interact. It is not a required hidden ledger, reasoning procedure or extra self-review pass; one focused test need not load it. Preserve explicit requested evidence rather than prescribing how to think.
Choose failures around the affected durable boundary: intent/acceptance loss, effect-before-ACK, ACK loss, competing attempt, ownership expiry, poison/retry, replay or mixed-version behavior where those paths matter. A full readiness claim must cover both sides of each relevant commit. A new quarantine guarantee includes bounded redrive preserving original identity and business state, not just a DLQ.
Pin engine/client, topology, durability settings and relevant workload. Assert business state plus actual broker progress/identity for broker claims; a once-called mock handler is insufficient. Use existing local harnesses and a test that would distinguish the absent commit/idempotency mechanism. Explicit runnable requests need source, setup and assertions; design/audit evidence may be proposed or unrun. Unavailable infrastructure is a disclosed limit, not an inferred guarantee.
Report and complete
Lead with verdict, affected guarantee, cause/design, recovery, observed proof and authority/gaps. Keep all requested artifacts and distinguish proposed, locally implemented, tested, deployed and verified-live states. A local fix can be complete without claiming production readiness. Reuse valid same-revision/environment results, repair introduced failures and stop when the requested result and applicable checks are satisfied or a concrete blocker is reported. Do not invent new clusters, full load campaigns or unrelated cleanup to finish a narrow task.
Adjacent job execution, schema, external API and cache decisions may use their matching skills when genuinely unresolved. Names are navigation, not dependencies or automatic handoffs; preserve this skill's identity and delivery ownership.