On this page

A server can be online while a customer’s reward journey is stuck. The application may accept requests successfully but leave orders unevaluated, wait indefinitely for a provider outcome or display a status that has not refreshed. Infrastructure uptime is useful evidence, but it is not a complete measure of the integration’s health.

A hypothetical merchant sees green server checks while customers ask why their reward references remain pending. The engineering team needs a scorecard tied to the stages customers depend on, with each signal pointing to a specific investigation rather than an undifferentiated “system unhealthy” alarm.

Define a measurable healthy customer journey

Describe health as observable progress through the approved workflow. An authorized purchase event is accepted, required inputs are established, the permitted evaluation completes and the customer or staff view receives a sufficiently current projection. Provider-controlled stages remain separate and use the actual contract’s evidence.

Do not define health as “every order receives a card.” Some purchases will not qualify, some require review and some remain pending for legitimate program reasons. A healthy integration records those outcomes accurately and moves each case only when the approved next step is available.

For the fictional merchant, a proposed scorecard is:

Signal Definition Source Response owner
Intake failure Verified delivery cannot enter durable processing Intake and platform diagnostics Integration operations
Processing lag Age of eligible queued work awaiting the next technical step Queue and stage records Worker service owner
Input uncertainty Required fields remain unresolved Normalization results Source adapter owner
Unknown external outcome Sent operation lacks an established result Operation ledger Provider integration owner
Stale display Customer projection is older than its approved freshness target Projection metadata Display service owner
Recovery gap Source comparison finds omitted processing Bounded reconciliation job Synchronization owner

These are proposed internal signals. They are not a list of telemetry guaranteed to be available from RebateCardX.

Shopify’s webhook troubleshooting tools can help establish delivery failures and performance evidence. Combine that source with local processing stages rather than treating platform delivery success as proof that the reward workflow finished. Shopify webhook troubleshooting

Define the denominator for every rate. “Five percent failed” is ambiguous unless the team knows whether it describes deliveries, unique orders, evaluation attempts or provider operations. A repeated technical attempt should not inflate the apparent number of affected customers without an explicit reason.

Track processing lag and unresolved handoffs separately

Measure lag for work that is actually ready to proceed. A purchase waiting under the campaign’s approved timing rule is not the same as a worker backlog. If both appear in one pending queue, the scorecard can generate noisy alerts or conceal a real technical delay.

For the hypothetical merchant, a queued evaluation may be technically ready but unprocessed for twenty minutes, while another reward is correctly waiting for a later purchase event. The first belongs in worker lag; the second belongs in a business-stage population. Neither should be labeled from age alone without its readiness context.

Unknown external outcomes deserve a separate signal. A provider timeout after submission requires outcome resolution. Counting it as an ordinary failed job can encourage retries that repeat a consequential action. The scorecard should show the number and age of unresolved operations, their last-established evidence and the owner responsible for resolving them.

A useful internal state split is:

pending_because_program_condition_is_not_yet_met
ready_but_waiting_for_technical_processing
blocked_by_missing_required_input
waiting_for_known_provider_processing
external_outcome_unknown_after_attempt
complete_but_customer_projection_stale

These labels are original design suggestions, not provider API statuses. Their purpose is to keep distinct problems from collapsing into one pending count.

Measure distribution as well as averages. A low average processing time can hide a small group of very old cases. The oldest ready task and selected latency percentiles can help identify that tail, provided the team defines the population and timing boundaries consistently.

A difficult case is a campaign with very low volume. A rate can swing dramatically after one event, while an absolute count alone may look unimportant. Use a combination of affected-case count, age and workflow consequence rather than importing alert thresholds from a high-volume merchant.

Segment by trusted merchant and integration version where useful. A single store with a changed permission can experience failures while the global average remains healthy. The scorecard should allow the owner to locate that scope without exposing unrelated merchants’ data.

Choose thresholds from the actual service objective and observed workflow, then label them as internal targets unless they are part of a confirmed contract. Do not present illustrative numbers as a provider SLA.

For example, the fictional team might investigate when the oldest technically ready evaluation exceeds its internally chosen target, or when any consequential operation remains unknown beyond the approved resolution interval. The exact intervals require the real contract and operating context. The valuable part is the action attached to the signal.

An alert definition should contain:

Element Filled illustrative example
Signal Ready evaluation backlog age
Scope One trusted merchant and worker version
Trigger Reviewed internal age threshold exceeded
First question Are workers consuming this queue and obtaining required source data?
Evidence link Safe queue and operation references
Owner Worker service operator
Escalation condition Backlog continues after the identified repair or source access remains blocked
Customer implication Evaluation updates may be delayed; no assumption about already issued cards

The last row prevents overbroad incident messaging. A stuck evaluation queue does not automatically mean existing cards cannot be used. The team should communicate the affected task, not infer failures in unrelated provider services.

Test alerts with synthetic failure states. Pause a test worker, withhold a required source field, simulate an unknown provider outcome and freeze a customer projection. Each scenario should trigger the intended signal and owner. If all four produce the same alert, the scorecard is not yet helping the team distinguish the problem.

Also test recovery. After the fixture is resolved, the signal should clear only when the underlying state is established, not merely when the server restarts. A restart can restore infrastructure while leaving the original jobs unprocessed.

Keep logs and evidence links free of card credentials and live claim tokens. The owner needs safe correlation references and stage information, not a full customer payload in an alert sent to a broad channel.

Retain a filled scorecard review showing one healthy period and one synthetic incident. Explain why the chosen signals changed and which action restored the workflow. This verifies that the scorecard responds to customer-impacting state rather than just resource utilization.

The finished monitoring design connects measurable stages to accountable investigation. It lets the team distinguish healthy waiting from broken processing and react to the actual blocked customer task while the servers may still be reporting perfect uptime.

Include a quiet-period fixture where there are no new purchases. Zero intake should not trigger the same alarm as failed intake when no work was expected. Then add one synthetic ready task and stop its worker. The scorecard should now identify a blocked task even though the overall volume remains tiny. This pair tests whether the monitoring model understands expected work rather than assuming that a busy event stream is always healthy and a silent stream is always broken. Retain the expected population and readiness state with the alert evidence.

Request technical access to discuss the observability available for your integration.

Request technical access

Source references

Back to contents

General information only

This guide is general information, not financial, legal, tax or regulatory advice. Eligibility, card availability, permitted use and responsibilities depend on the applicable offer and card terms.