Monitor integration reliability through completed business tasks
Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.
Your integration dashboard shows successful requests, but staff still chase missing records or discover that updates arrived after their useful deadline. The dashboard may be accurate about requests while answering the wrong operational question.
Measure reliability around a stable business task: what work was accepted, what proves completion, when that completion occurred, and what remains unresolved. Keep request success as a diagnostic indicator, not a substitute for completed work.
The fictional worked example below provides a reusable scorecard with explicit denominators, evidence coverage and recovery actions. It is a recommended operating design, not a claim about controls available in a particular product.
1. Define one business task and its acceptance boundary
A task is one intended business outcome, not one API call, queue message or retry. Write its contract before choosing dashboard metrics:
- Acceptance: when the service durably records responsibility for eligible work.
- Identity: a stable task identifier linking the source change, execution attempts and destination evidence.
- Completion: the business condition that must be verified, including correctness checks.
- Deadline: when completion stops being timely.
- Evidence: the records that establish completion and its timing.
- Owner: who acts when the outcome cannot be established or completed.
For this hypothetical workflow, the contract is: make one accepted source update visible in the destination, with the expected source version, required fields and owner, within 15 minutes of acceptance.
A transport acknowledgement does not satisfy that contract. Neither does a destination record containing an older source version. Evidence must match the intended result.
Do not record acceptance only after successful processing; that would hide lost work from the denominator. Track upstream submissions and pre-acceptance rejections separately so a strong completion rate cannot conceal an intake failure.
2. Separate task identity, execution attempts and outcome states
Maintain two linked ledgers. The task ledger holds the task ID, acceptance time, deadline, intended source version, outcome, completion evidence, timing evidence and recovery owner. The attempt ledger holds each execution attempt and its diagnostic result.
Retries remain attached to the original task. A genuinely new source change receives a new task identity, even if it affects the same destination record.
For a cohort whose deadlines have passed, use five mutually exclusive states:
- Completed on time: verified completion occurred at or before the deadline.
- Completed late: verified completion occurred after the deadline but by the reporting cutoff.
- Completed, timing uncertain: completion by the cutoff is verified, but the evidence cannot establish which side of the deadline it occurred on.
- Known unresolved failure: evidence establishes a failure or missing required result as of the cutoff, without verified completion.
- Unknown outcome: evidence cannot establish either completion or a known unresolved failure.
The timing-uncertain state matters because uncertainty about when work finished differs from uncertainty about whether it finished. Neither should be quietly counted as timely success.
3. Distinguish completion time from observation time
In this fictional workflow, an authoritative destination commit record normally establishes completion time. It identifies the task and source version, confirms the required fields and owner, and records when the intended update became available. Comparing that timestamp with the stored deadline establishes timely or late completion. The design assumes comparable clocks; if that assumption is unreliable near a deadline, timing remains uncertain.
Some tasks lack an authoritative completion timestamp. Polling may still prove completion, but a poll timestamp is not the actual completion time.
For example, a hypothetical task accepted at 09:20 has a 09:35 deadline. An authoritative check at 09:34 finds the intended version absent; another at 09:37 finds it present. Completion occurred after 09:34 and no later than 09:37. Because that interval straddles 09:35, classify it as completed with uncertain timing—not late merely because the successful poll arrived at 09:37.
This inference requires trustworthy observations and a result that cannot appear and disappear between checks without detection. Otherwise, retain an unknown outcome until stronger evidence is available. When a reliable completion interval falls entirely before or entirely after the deadline, it can establish timely or late completion without an exact timestamp.
4. Fix the cohort, reporting cutoff and evidence watermark
Use a matured acceptance cohort: tasks accepted during a defined interval whose deadlines have all passed by the reporting cutoff. Otherwise, newly accepted work can make the completion rate look worse simply because it has not had time to finish.
All example times, counts, limits and targets in this article are hypothetical. The worked cohort contains 1,000 distinct tasks accepted from 09:00 up to, but not including, 10:00, each with a 15-minute deadline. There are no cancellations. All times use the same reporting time standard.
- Reporting cutoff: 10:30. Outcomes describe the cohort as of this time.
- Last successful reconciliation run: 10:31.
- Evidence watermark: 10:30. Acceptance, destination completion, failure and reconciliation evidence is fully processed through this time.
- Publication time: 10:31.
- Freshness limit: a hypothetical maximum two-minute lag between publication and the evidence watermark.
The one-minute lag meets that limit, and the watermark covers the reporting cutoff. These are separate checks. A reconciliation that ran at 10:29 would not, by itself, establish coverage through 10:30.
A watermark must reflect confirmed processing coverage across every evidence stream needed for the scorecard—not simply the newest timestamp observed. Here, reconciliation also checks destination state as of the cutoff. That supports the unresolved counts, while individual exceptions can still leave some outcomes unknown.
If coverage reached only 10:29, either report through that covered cutoff or label the 10:30 results provisional with a one-minute coverage gap. Completion rates and unresolved counts would both need that qualification: work might have completed during the uncovered interval.
5. Reuse this filled reliability scorecard
The following artifact is filled for the fictional cohort. Targets are internal operating choices, not benchmarks, contractual promises or Alactic results. Every task rate uses the same 1,000 accepted-task denominator.
| Field | Filled value or rule |
|---|---|
| Task contract | Expected source version, required fields and owner visible in destination within 15 minutes of acceptance |
| Cohort | 1,000 distinct tasks accepted 09:00–10:00, excluding 10:00 |
| Reporting coverage | Cutoff 10:30; watermark 10:30; last successful run and publication 10:31 |
| Freshness rule | Maximum two-minute watermark lag; actual lag one minute; cutoff fully covered |
| Acceptance evidence | Durable acceptance ledger reconciled against intake records |
| Completion evidence | Destination verification of intended result, linked to authoritative commit record or reliable observation interval |
| Recovery owner | Integration operations lead; business process owner handles missed handoffs |
| Indicator | Calculation | Interpretation and action |
|---|---|---|
| Verified on-time completion | 850 / 1,000 = 85%; target at least 97% | Inspect processing delays and queue age. Another 20 completed tasks have uncertain timing; do not count them as on time. |
| Verified completion by cutoff | (850 + 50 + 20) / 1,000 = 92%; target at least 99% | Includes timely, late and timing-uncertain completions. Full evidence coverage through 10:30 supports this figure. |
| Known late completion share | 50 / 1,000 = 5% | Inspect missed deadlines and any business follow-up still needed. |
| Completed, timing uncertain | 20 / 1,000 = 2% | All 20 observation intervals straddle their deadlines. Seek authoritative timing evidence; otherwise retain the classification. |
| Known unresolved failure share | 30 / 1,000 = 3%; target at most 0.5% | Failure records and destination reconciliation establish unresolved status at 10:30. Assign cause correction and recovery. |
| Unknown outcome share | 50 / 1,000 = 5%; target at most 0.5% | Reconciliation identifies insufficient task-level evidence despite covered processing streams. Investigate before replay. |
The five states reconcile: 850 on time + 50 late + 20 completed with uncertain timing + 30 known unresolved + 50 unknown = 1,000 accepted tasks. The 92% completion indicator combines the first three states; it is not another state to add.
The verified on-time figure is 85%, not proof that exactly 15% finished late. Even if all 20 timing-uncertain completions were later shown to be timely, those tasks alone would raise it only to 87%. The 80 unresolved or unknown tasks remain separate.
Store evidence references alongside these totals so reviewers can inspect classifications. Apply the declared freshness limit to acceptance, completion, failure and exception evidence, not merely to the dashboard refresh.
6. Explain why 99% request success can coexist with 85% timely completion
In the same fictional example, monitoring records 5,000 outbound request attempts associated with the cohort. Of these, 4,950 return responses classified as transport successes: a 99% request-success rate.
Those attempts include writes, status checks and retries. Successful checks can repeatedly report pending processing. A write can be acknowledged before downstream validation finishes. Consequently, request success does not establish task completion.
The business result remains 850 verified timely completions out of 1,000 accepted tasks. Another 50 completed late, 20 completed with uncertain timing, 30 have known unresolved failures and 50 have unknown outcomes.
Keep the request metric because it helps diagnose transport behaviour. Label its unit and denominator, and never average it with task completion. Compare problem cohorts by task type, destination and acceptance period; avoid combining workflows with different deadlines or definitions of success.
7. Recover according to the evidence, not a blanket retry rule
A single retry action is not a recovery policy. Each state requires a different decision:
- Known unresolved failure: correct the cause through the authorised process, then assess replay. Preserve task identity and the original deadline.
- Unknown outcome: reconcile the destination first. A timed-out write may already have taken effect. Repeat execution only when authorised and safe; otherwise escalate with the available evidence.
- Timing-uncertain completion: investigate timestamps or observation coverage. Do not re-execute completed work to resolve a measurement problem.
- Late completion: check whether staff still need to repair a missed handoff. Track that follow-up separately rather than rewriting the original result as timely.
- Stale evidence: restore coverage and reconcile the missing interval before presenting current totals as final.
Rate limiting deserves a separate diagnostic category. HTTP 429 indicates rate limiting; a response may include a Retry-After header. Providers apply limits differently. Respect available retry guidance without treating a retry as a new allowance of business time.
Later evidence may legitimately revise a historical classification. Retain the earlier snapshot, record the reason and publish the revised result. Recovery today must not erase yesterday's lateness.
8. Pilot one task type before expanding
Start with one workflow whose missing or late results staff already chase. Agree its acceptance boundary, completion evidence, timing rules, deadline and recovery owner. Run a daily matured cohort for a hypothetical five-working-day pilot.
Require three observable checks at each review: every accepted task belongs to exactly one state; every unresolved or uncertain case has an owner and next action; and evidence coverage reaches the stated reporting cutoff within the freshness limit.
Compare staff-reported stuck work with the ledger. A task staff can identify but monitoring cannot account for exposes a concrete gap in acceptance capture, identity or evidence. Fix that gap before expanding.
Verification also has a cost: frequent checks increase load, while sparse checks widen timing uncertainty and delay recovery. Choose a cadence that supports useful decisions. If authoritative timestamps are unavailable and polling cannot resolve deadline performance, publish that limitation rather than presenting false precision.
Put the next conversation to work.
Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.
Explore Communications