A technical due diligence review of background processing.
Request evidence that accepted jobs reach correct business outcomes after crashes and repeated delivery. Distinguish demonstrated defects, untested assumptions, and operational limits, then give important findings a reproducible acceptance test.
Inside this article 6 sections +
Technical due diligence for background processing should establish what happens to accepted work when delivery repeats, workers fail, and dependencies remain unavailable. Request evidence of business completion and recovery, not just a diagram naming a queue technology.
The review should distinguish demonstrated behavior, plausible design, and unresolved risk. A queue can be professionally operated while application handlers still duplicate charges or lose accepted work. Conversely, a simple queue design may be appropriate when its limits are explicit and tested.
Start with the business work in the queue
Inventory the job classes and their consequences. Generating a disposable thumbnail, issuing a customer refund, updating search results, and sending an expiry reminder require different evidence. Identify the durable source of intent, the expected completion deadline, and the external effects for each class.
Choose a representative high-consequence workflow for a detailed trace. In an illustrative subscription platform, use account credit issuance. Follow the request from authorization through durable acceptance, enqueueing, execution, provider response, and customer-visible completion. Record every point where the system can fail between two writes.
Ask who owns unresolved work. A dead-letter queue with no review process is a storage location, not an operating control. A dashboard showing job counts does not establish that the team can explain which customer operations are incomplete or safely repair them.
Request a small evidence pack
Ask for queue configuration, handler code for the selected workflow, deployment shutdown behavior, retry policy, dead-letter handling, and the operation-state model. Include an example incident or controlled recovery exercise with sensitive data removed. Request read-only access or bounded extracts appropriate to the review scope.
For SQS, inspect queue type, visibility timeout, message retention, redrive configuration, and consumer behavior. Standard queues use at-least-once delivery, so application correctness must tolerate repeated delivery. A FIFO configuration does not remove the need to reason about effects committed before acknowledgment and retries outside transport deduplication behavior.
For BullMQ, inspect Redis durability assumptions, lock renewal, stalled-job handling, retries, and application idempotency. The meaningful question is the same across technologies: what evidence connects a customer request to one correct business outcome after interruption?
Review three failure windows
The first window is between saving business intent and enqueueing work. If those writes are separate, a crash can leave accepted work unscheduled. Look for an outbox, reconciliation sweep, or another explicit mechanism that closes the gap. Confirm that the recovery mechanism itself has an owner and an alert.
The second window is between performing an external effect and recording completion. A provider may accept a credit while the worker times out. Look for a stable operation key, provider result lookup, or a reconciliation process. A generic retry that assumes every timeout means “nothing happened” is a material finding.
The third window is between completion and queue acknowledgment. The job may be delivered again after its business result already exists. Inspect whether the handler recognizes that result and returns safely. Deleting a queue message is not the only evidence of completion, and retaining a message is not proof that the effect remains undone.
Test recovery with bounded exercises
Use a sandbox workflow and test recipient or provider account. Interrupt the worker at each window, restart it, and inspect the operation record and external effect. The expected result should be written before execution. Do not use real customer transactions as a convenient diligence fixture.
Delay the dependency long enough to exceed ordinary processing time. Observe visibility or lock renewal, overlapping deliveries, retry amplification, and backlog age. Then restore the dependency and verify recovery throughput while normal traffic continues. A system that survives the outage but cannot drain its backlog still has an operating limit.
Test deployment interruption separately from an abrupt process kill. Graceful shutdown may cover routine releases, while crashes require durable recovery. The job reconciliation guide provides evidence for finding missing or ambiguous operations after either event.
Prioritize findings by consequence and evidence
Describe each finding as a concrete failure condition, customer consequence, supporting evidence, and proposed next action. “Queue architecture is immature” is not useful. “A timeout after provider acceptance can issue the same credit twice because retries create a new provider key” identifies a defect that can be reproduced and repaired.
Separate a demonstrated correctness defect from an untested recovery assumption or an operational efficiency opportunity. They require different urgency and confidence. Do not collapse them into a single percentage score that suggests more precision than the review supports.
Record compensating controls and their limits. A daily manual reconciliation may reduce exposure, but only if it covers the relevant operations and has a reliable owner. A small workload may make a manual process viable today; growth assumptions should identify when that control becomes insufficient.
Leave a review that supports a decision
The output should state which workflows were inspected, which exercises ran, which evidence was unavailable, and which conclusions remain conditional. Include prioritized remediation with an acceptance test for each important finding. Avoid implying that a bounded sample certifies every queue and handler in the estate.
Connect findings to the commercial or engineering decision: whether to proceed with an integration, where to invest before increasing volume, or which operating capability a new team must own. The architecture review guide supplies the broader context. Technical consulting can turn queue evidence into a practical risk and remediation discussion without substituting technology preference for demonstrated behavior.