The short version
A dead letter queue holds messages that could not be delivered or processed, so a team can investigate them separately. For AI-assisted business workflows, recovery should classify each failed task, check whether its intended action is still valid, and assign an owner before returning eligible work for another attempt.
What is a dead letter queue?
Microsoft's Azure Service Bus documentation describes the dead-letter queue (DLQ) as a place to hold messages that can't be delivered or processed, so they can be inspected. That makes it useful when an agent-driven workflow depends on background processing. It does not decide whether the original business request should still happen. That decision belongs in the application and its operating procedure.
Illustration, not a customer case: an AI assistant prepares equipment-service requests for a legacy maintenance system. Some background tasks fail. While engineers investigate, an operator completes one request manually and a customer cancels another. Emptying the failed-message queue indiscriminately could now recreate work that should not be repeated.
How to handle a dead letter queue: classify the work
We recommend an operational record for each failed business task: the stable request identifier, failure category, current owner, last verified business state and next decision. Keep sensitive payloads behind appropriate access controls. A dashboard should explain what needs attention without copying entire customer messages into every alert.
Classify the illustrative backlog into four decisions. A temporary dependency failure may become eligible after recovery. Invalid input needs correction. An unknown remote outcome needs reconciliation. A canceled or already completed request needs a recorded closure decision. The same technical error code can hide different business situations, so do not let the agent infer eligibility from an exception message alone.
- Retry candidate: the cause is addressed and the intended operation remains authorized.
- Correction required: repair or regenerate the proposed input, preserving its relationship to the original request.
- Investigation required: check the authoritative system before deciding whether another attempt is safe.
- No further execution: record why the request is obsolete or already satisfied.
Dead letter queue redrive is a controlled release
Amazon SQS documentation, checked on 18 November 2026, describes redrive as moving messages from a DLQ to another queue for processing. Its managed redrive does not filter or modify messages, and redriven messages receive new message identifiers. Consequently, our recommendation is to preserve a business request identifier independently of the broker identifier. Do not treat redrive as an editor that will fix a mixed backlog.
Before enabling recovery, define how eligible work is selected and how consumers handle ineligible records if they encounter them. A revised input should have traceable review history. Do not silently replace the original evidence with a new model answer. For the maintenance illustration, the operator should be able to see which request was corrected, why and which version was executed.
Release eligible work gradually and agree on stop conditions: repeated validation failures, excessive downstream delay or unexpected duplicate business records. SQS lets you set a custom maximum redrive velocity and recommends starting small, but the safe operating rate remains a decision for your system. Measure the legacy dependency and current workload rather than choosing a rate from the queue's maximum capacity.
An empty queue is not proof of recovery
Retention behavior is provider-specific. AWS documents expiration based on the original enqueue time for standard-queue messages moved into a DLQ. Azure Service Bus documents no automatic DLQ cleanup. Check the actual service configuration and preserve the business investigation record separately, so disappearance or indefinite retention does not become your definition of resolution.
For the illustrative workflow, rehearse recovery with synthetic requests in a test environment. Include a corrected task, an already completed request, a cancellation and an unresolved remote result. Verify business records after processing and ensure each request has an explicit disposition. This is a proposed test plan, not a test result.
Track unresolved business-task age, ownerless items, repeat failures and verified completions. AI can help group errors and prepare investigation summaries, while engineers retain responsibility for eligibility rules and recovery controls. UVISION provides dedicated developers and managed projects with at least one full-time developer per project, maintaining critical legacy systems and evolving AI workflows as part of long-term engineering ownership.
Common questions
What is a dead letter queue?
A dead letter queue holds messages that could not be delivered or processed, allowing separate investigation. In an AI-assisted workflow, it helps expose failed background work. It does not determine whether the business action is still appropriate. Assign an owner and check current request state before deciding how each task should be resolved.
How to handle a dead letter queue?
Classify failed tasks into retry candidates, corrections, investigations and closures. Keep a stable business request identifier and evidence of the latest verified state. In the maintenance illustration, a request completed manually should be closed appropriately rather than executed again. Track unresolved age and ownership instead of using queue size as the sole recovery measure.
What is dead letter queue redrive?
Redrive moves failed messages back toward processing. It should follow an eligibility decision, not substitute for one. For the illustrative maintenance workflow, release eligible requests gradually and stop when repeated failures or unexpected duplicates appear. Verify the resulting business records, because moving a message does not prove that the intended work was completed.
Further reading
Want to explore what this means for your systems?
Explore the relevant service