Service-level objectives for asynchronous customer work.
Measure correct completion within the promised interval from durable acceptance. Include overdue unfinished work and evaluate mature cohorts so missing or slow jobs cannot disappear from the denominator.
Inside this article 6 sections +
An asynchronous workflow service-level objective should measure whether accepted customer work reaches a correct result within its promised time. Enqueue success is an acceptance signal, not proof of completion. Count unresolved work that has passed its deadline, even when no failure event was emitted.
Define the population and clock before building an alert. A dashboard based only on completed jobs can improve during an outage because the slowest jobs disappear from its denominator.
Choose the customer promise and acceptance boundary
For an illustrative data-export feature, the promise is that a valid, accepted export becomes available within five minutes. Acceptance occurs when the application durably records the operation and takes responsibility for processing it. A request rejected before that boundary is a separate API outcome.
Define what “available” means: the file exists, its manifest is complete, the customer has permission to retrieve it, and the operation status exposes the result. A worker marking itself finished before the file is readable does not satisfy the customer's promise.
Separate correctness from timeliness. An export produced quickly with missing rows is not good service. The completion record should include validation evidence sufficient for the workflow, while the SLO uses a clear rule for which outcomes qualify.
| Operation | Accepted | Observed result | Five-minute cohort decision |
|---|---|---|---|
| A | 0 | Completed at 4 | Good |
| B | 1 | Completed at 7 | Late; bad |
| C | 3 | Still pending | Overdue; bad |
| D | 6 | Still pending | Not yet mature |
| E | 2 | Terminal failure at 3 | Bad |
Use acceptance cohorts to avoid survivor bias
Group operations by when responsibility was accepted. Evaluate a cohort after its completion window has elapsed, or expose provisional results with clear maturity rules. Do not compare a fully observed historical cohort with a current cohort that still has time to finish.
The following Python fixture implements only the table's timing rule. Production calculation also needs a durable operation population, correctness status, cancellation policy, and evidence completeness checks.
rows = [
("A", 0, 4, "completed"),
("B", 1, 7, "completed"),
("C", 3, None, "pending"),
("D", 6, None, "pending"),
("E", 2, 3, "failed"),
]
observed_at, deadline = 10, 5
mature = [r for r in rows if r[1] + deadline <= observed_at]
good = [r for r in mature
if r[3] == "completed" and r[2] - r[1] <= deadline]
assert len(mature) == 4
assert len(good) / len(mature) == 0.25
Retain a separate view of current pending age. Mature-cohort reporting is useful for the objective, while an oldest-unfinished-work signal can reveal a developing incident before the reporting period closes.
Define failures, cancellations, and missing evidence
A terminal processing failure after acceptance normally fails the completion promise. A user-requested cancellation may have a different policy if the product explicitly allows it. Record that policy before an incident so failed work cannot be relabeled as cancelled to improve the metric.
Use durable operation state to identify missing completions. A telemetry stream can lose the final event, and a worker can disappear without emitting an error. Reconciliation should distinguish an incomplete operation from missing monitoring evidence, while both conditions remain visible to operators.
Specify duplicate handling. Several deliveries of one job represent attempts at one accepted operation, not several customer successes. The job reconciliation guide explains how operation identity supports that accounting.
Separate workload classes before combining the promise
A small interactive export and a large archival export may have different completion promises. Record the class at acceptance using a stable rule, then evaluate each operation against the deadline that applied when it was accepted. Reclassifying overdue work into a slower category after the fact would hide a missed promise.
Publish both class-level results and any combined view. A high-volume fast class can dominate the aggregate while a smaller paid capability remains unreliable. If the business chooses a combined objective, document how the populations are weighted and retain the separate operating signals.
Monitor the evidence pipeline itself. Reconcile the accepted-operation count against the durable source and inspect gaps in status updates. A missing batch of telemetry should produce an unknown or incomplete reporting state, not an apparently perfect success rate.
Connect the objective to operating decisions
Choose the target with product and operations owners based on customer use. An illustrative 99% target over 1,000 eligible operations allows ten bad outcomes in that population. It is a planning example, not a recommendation that every export service should use that target.
Define what happens when the error budget is consumed or threatened. The response might pause risky changes to that workflow, prioritize a dependency repair, or reduce accepted load until recovery capacity is restored. A metric without an agreed decision becomes a report that nobody owns.
Keep low-volume workflows in view. A ratio based on a handful of operations can move sharply, and a single important customer's overdue work may deserve action even when a fleet-wide percentage looks acceptable. Combine the SLO with account-aware operating signals where the product requires them.
Validate the metric before alerting on it
Run fixtures for timely completion, late completion, terminal failure, duplicate delivery, missing completion evidence, and an immature cohort. Test boundary times exactly at the deadline. Confirm that a failed ingestion of operation records cannot make the denominator shrink silently.
Then connect alerting to urgency and recovery action. The customer-impact alert review helps separate immediate intervention from a slower reliability trend. Technical consulting can align the workflow promise, measurement, and operating response so the objective supports real service decisions.