The Queue That Remembers Nothing: Auditing Undelivered Events Before the 2026 Close
Photo: server room message queue monitoring dashboard operations engineer, via assets.gcore.pro
Most UK integration estates are built to succeed loudly and fail quietly. A message is produced, accepted by the broker, handed to a consumer, and there the visible story usually ends. When the consumer rejects it — a schema drift, a downstream timeout, a duplicate key — the platform does exactly what it was configured to do: it retries a defined number of times, then parks the message in a dead-letter queue. The queue then sits there, holding the only record of a business event that never landed.
Nothing on the executive dashboard turns red. The order status screen shows the last state it knew about. The reconciliation job runs against a marginally smaller population than it should. Weeks later somebody notices a customer billed but never provisioned, or a stock adjustment that never reached the warehouse system, and the investigation begins at entirely the wrong end of the chain.
Why silence is the default setting
Dead-letter queues are almost always created by default and governed by exception. The infrastructure team owns the broker, the application team owns the consumer, and the business outcome — the invoice raised, the appointment booked, the claim advanced — belongs to nobody in particular. Alerting, where it exists, tends to fire on queue depth rather than message age, which means a single stuck message on a Monday morning looks identical to a scheduled batch arrival on a Sunday night.
There is a second, subtler failure. An empty dead-letter queue is not proof of delivery. It is proof that nothing was rejected in a way that the consumer recognised as a rejection. Silent acknowledgements, catch-and-log exception handlers and retry policies that swallow the final attempt all produce a clean queue and a broken process.
What the queue does not tell you
The dead-letter queue preserves payloads. It rarely preserves meaning. Headers are stripped by an API gateway, correlation identifiers are regenerated at each hop, timestamps arrive in UTC while the operational team works in local time, and the business key that would connect the message to a customer record may sit three levels deep in a JSON body that no analyst has been trained to read.
Retention is the other constraint. Managed brokers commonly set short default retention windows, and some configurations purge on redeploy. If retention is shorter than the longest reconciliation cycle in your estate, then for part of every reporting period you cannot prove delivery at all.
Four questions to answer before 31 December
Do we have a complete inventory?
Start with the obvious queues and then look for the ones nobody documented: topics in a secondary region, queues created by a supplier during onboarding, the subscription spun up for a pilot that quietly became load-bearing. A register that lists only what the platform team knows about will miss precisely the flows most likely to fail.
Can we attribute a message to a process?
Take a sample — twenty messages from each queue that has carried traffic in the last ninety days — and ask a business analyst to name the customer, the process and the financial consequence of the message not arriving. If they cannot, the queue is an engineering artefact rather than an operational one, and it needs a translation layer before it can be governed.
Has anyone ever replayed, and is replay safe?
Replay is a procedure, not a button. Idempotency at the consumer, ordering guarantees, duplicate suppression downstream and partial replay of a subset of messages all have to be understood before anyone presses anything. A replay that works in the test estate and doubles a customer's credit in production is worse than no replay capability at all.
What does the alerting actually measure?
On depth, a healthy high-volume queue and a stalled low-volume queue look similar. On age, they do not. Age-based alerting on the oldest unprocessed message per flow is a cheap and disproportionately effective control.
The close-period exposure
The weeks around year-end compress the margin for error. Peak trading volumes, payroll cut-offs, statutory returns and the December change freeze all arrive together, and some teams reduce alert noise during a freeze by suppressing notifications altogether — a reasonable-sounding decision with a well-documented failure mode.
A missing event in February is an anomaly. A missing event in the week before close is a restatement risk, and it will be discovered by an auditor rather than by you.
Proving delivery rather than absence
The audit's real output is not an emptied queue. It is a small set of controls that make absence visible: produced-versus-processed counts reconciled daily for each critical flow, a canary event generated end-to-end each morning and checked against the downstream table, and freshness contracts on the systems of record that the events feed.
A workable control set is modest. A register of every dead-letter queue with a named business owner alongside a platform custodian. Age-based alerting rather than depth-based. A documented and rehearsed replay procedure, exercised at least twice a year. Retention configured to exceed the longest reconciliation cycle in the estate. And a monthly review in which a human being reads a sample of payloads and confirms that each one either arrived somewhere or has a reasoned exception against it.
Operations teams tend to treat the dead-letter queue as a technical bin. It is better understood as the ledger of everything the business tried to do and the systems declined to complete. Before the 2026 close, it is worth reading the ledger.