Back to all articlesIntegration reliability

Choose a safe degraded mode when a dependency is unavailable

Alactic Editorial10 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

When an inventory dependency becomes unavailable, stopping every customer conversation may be unnecessary. Continuing to promise stock availability may be worse. The useful choice is not whether the entire workflow stays open: it is which actions remain safe without the missing information.

Choose a degraded mode for each action using three tests: can its consequences be reversed, how fresh must its inputs be, and will the person using the fallback understand what has and has not happened? Stop actions that need authoritative information you cannot obtain. Keep safe information visible in read-only mode. Defer requests whose execution can wait. Use older data only for explicitly bounded purposes.

The decision table below turns those tests into operating instructions. It also separates two questions that often get confused: whether an action is permitted during an outage, and whether the organisation has a verified way to support it.

1. Start with the decision, not the unavailable system

A dependency can support several decisions with very different consequences. An unavailable stock service might prevent a reliable reservation while leaving product descriptions useful. An unavailable scheduling service might prevent confirming an appointment while still allowing staff to record a customer's preferred time.

List the actions affected by the outage. For each, identify the missing fact, the consequence of being wrong, and the latest point at which the action can safely wait.

  • Reading: Is staff merely viewing information, or will the customer reasonably treat that information as a promise?
  • Recording: Is this a note of interest, a request awaiting review, or a commitment?
  • Changing: Does the action reserve capacity, change ownership, spend money, or trigger another process?
  • Communicating: Does the message accurately describe the current state, including any uncertainty?

Do not classify an action as harmless just because it changes no database record. A staff member reading an old availability count aloud can create expectations that are difficult to unwind. Evaluate the business effect, not only the technical write.

2. Apply three tests before allowing continuation

Reversibility: can you undo the consequence?

A draft note is usually easier to correct than a confirmed commitment. But reversibility includes the customer's effort, staff work, and downstream changes—not just whether a record can be edited. If undoing an action requires contacting customers or recovering allocated capacity, treat it as consequential.

For each permitted action, name its correction path. If nobody can explain how to reverse an incorrect result, stopping that action is the safer default until an authorised owner decides otherwise.

Freshness: how old can the evidence be?

Different fields need different freshness rules. A product description may remain useful while a stock count becomes unsuitable for commitment-making. Define the permitted purpose and maximum age together. “Old data is allowed” is not an adequate rule.

Age alone is also insufficient. A record refreshed recently may still be incomplete or from the wrong scope. Require a known source, a last-successful-refresh time, and enough context to judge whether the information applies. If age or coverage cannot be established, treat freshness as unknown.

Understanding: will users interpret the state correctly?

A generic outage banner cannot explain every record. Put uncertainty beside the affected information and the action it limits. Distinguish request recorded, awaiting validation, confirmed, and failed. A reassuring completion message is unsafe if it merely means that a request entered a holding list.

If staff cannot reliably explain the fallback, reduce its scope. A more restrictive mode with clear instructions is preferable to a flexible mode that invites accidental promises.

3. A filled hypothetical degraded-mode decision table

Fictional worked example: A business equipment supplier loses access to its authoritative inventory service. Staff can still access some customer records and product information. The incident owner must decide how to handle stock questions and reservation requests.

All thresholds, labels, roles, and actions in this example are hypothetical operating choices, not Alactic capabilities or customer results. They assume the organisation has verified the necessary access and recordkeeping arrangements. If a required arrangement does not exist, the corresponding fallback is unavailable.

The following decision table is shown as row-by-row records so each instruction remains readable. Each row contains the same fields and can be copied into an incident runbook.

Row A — Stop stock commitments

  • Action covered: Confirming stock, reserving units, or promising a dispatch date dependent on stock.
  • Permitted actions: Explain the interruption and record a request for later review, if the separate deferred-request process is available.
  • Prohibited actions: Confirm availability, claim a reservation exists, or turn an old count into a dispatch promise.
  • Freshness label: “Live availability unavailable. Stock commitments paused.”
  • Staff instructions: Tell the customer that availability cannot currently be confirmed. Do not substitute a colleague's recollection for authoritative confirmation.
  • Exit criteria: The incident owner verifies authoritative inventory access, checks representative affected items, resolves uncertain earlier reservations, and authorises resumption.

Row B — Read-only customer context

  • Action covered: Viewing existing customer notes and previously recorded order references.
  • Permitted actions: Read records that remain authorised and verifiably unaffected by the failed dependency.
  • Prohibited actions: Edit affected order details, treat historical records as current stock evidence, or broaden access during the outage.
  • Freshness label: “Read-only customer context. Not evidence of current availability.”
  • Staff instructions: Use the information to understand the enquiry, not to confirm fulfilment. Route corrections through the designated follow-up process rather than creating competing copies.
  • Exit criteria: Relevant editing paths are checked, records awaiting correction have an owner, and the incident owner authorises normal editing.

Row C — Deferred reservation requests

  • Action covered: Recording a customer's request for a particular item and quantity.
  • Permitted actions: Record the request, its creation time, customer contact details, and an agreed contact preference in an approved holding process.
  • Prohibited actions: Describe the request as a reservation, promise queue position, collect duplicate requests as separate commitments, or execute requests without the required authorisation.
  • Freshness label: “Request recorded; stock not checked; nothing reserved.”
  • Staff instructions: Read back the requested item and quantity. Explain that staff will recheck availability before confirmation. Assign a review owner and preserve the request's status.
  • Exit criteria: Inventory access is verified; requests are checked for duplicates, cancellations, and current validity; an authorised person reviews each before execution.

Row D — Limited cached product reference

  • Action covered: Viewing a previously approved product description snapshot.
  • Permitted actions: Use descriptions and dimensions from a known snapshot no more than one business day old, solely for preliminary product discussions.
  • Prohibited actions: Use cached quantities, prices, or dispatch estimates to make commitments. Use a snapshot with unknown age or origin.
  • Freshness label: “Product reference snapshot: [actual refresh time and timezone]. Preliminary guidance only; availability not checked.”
  • Staff instructions: State that the description is from a snapshot. Refer uncertain specifications for later verification. Stop using it when the age limit is reached.
  • Exit criteria: Current authoritative product information is available and checked; outdated reference copies are withdrawn from the operating process.

These modes can coexist. The supplier stops stock commitments, allows read-only context, accepts carefully labelled requests, and limits older reference data. It does not select one mode for the whole business.

4. Make freshness a permission rule, not a footnote

A timestamp tells users when something was captured. It does not tell them what they may do with it. Pair the timestamp with an operational limit: “For product discussion only; not for availability confirmation.”

Define what happens when that limit expires. In the fictional example, an over-age product snapshot becomes unavailable for staff reference under this fallback; staff instead record the question for verification. The threshold is a local design choice that needs a business owner, not a universal standard.

Also define immediate disqualifiers. A known product change can make a snapshot unsuitable before its age limit. Missing records can make a seemingly current view misleading. An unknown refresh time should not be presented as recent.

Read-only and cached-data modes are different dimensions. Read-only restricts changes; it does not prove freshness. Cached-data mode restricts which older evidence may be used and for what purpose. A fallback may need both restrictions.

5. Treat deferred work as unresolved work

Deferred processing is useful only if the organisation can preserve intent without implying completion. Before accepting requests, establish an approved record, an owner, a capacity limit, and a way to communicate the unresolved state. If any of those are missing, stop accepting that category of request rather than build an invisible backlog.

A practical holding record includes the requested action, creation time, requester, current status, review owner, and any cancellation or expiry condition. Collect only the information needed for that purpose and use an approved location with appropriate access.

The dangerous recovery shortcut is to process everything as soon as the dependency responds. Conditions may have changed. A customer may have submitted again through another channel, withdrawn the request, or no longer need it.

  1. Identify requests that may already have succeeded before the outage became visible.
  2. Separate duplicates and cancellations from unresolved requests.
  3. Revalidate information that must be current at execution time.
  4. Obtain any required authorisation before acting.
  5. Record the outcome and communicate it accurately.

An uncertain earlier outcome is not the same as a failed action. Do not repeat it until its state has been investigated through an approved process.

6. Preserve access boundaries and give staff usable language

An outage does not justify granting broader access or copying sensitive records into an unrestricted shared document. If current permission cannot be established for an action or record, do not assume older access evidence is sufficient.

OWASP's authorization guidance recommends least privilege, deny-by-default behaviour, and permission checks on every request. These are design recommendations, not evidence that a particular product implements the fallback described here.

Staff also need language that matches the permitted action. In the fictional supplier example, a useful explanation is:

I can record the item and quantity you want, but I cannot check stock or reserve it right now. The request will need review before we can confirm anything.

That explanation sets a boundary without inventing a restoration time. Give staff an escalation route for urgent requests, disputed commitments, and unclear record states. Escalation should produce an authorised decision, not an informal exception that bypasses the missing check.

7. Exit by evidence, then test the decision table

A responding dependency is a recovery signal, not sufficient evidence that normal work is safe. Check whether affected records are current, earlier actions have known outcomes, and deferred work can be reviewed without creating duplicates or conflicting commitments.

Resume by action category where appropriate. Viewing current information may become safe before processing the backlog. Commitment-making may remain paused while uncertain reservations are reconciled. Keep a named incident owner responsible for recording which restrictions have been lifted and which remain.

Rollback conditions matter too. If checks reveal inconsistent availability, unresolved outcomes, or renewed failures, return the affected action to its previous restricted mode. Do not preserve a normal-looking workflow merely because recovery was already announced.

Next action: Take one dependency and list the decisions it supports. Complete the six fields used in the example for each chosen fallback: action covered, permitted actions, prohibited actions, freshness label, staff instructions, and exit criteria.

Then run a tabletop exercise with someone who did not write the instructions. Ask them to handle a fresh enquiry, an expired snapshot, a duplicate request, and an uncertain earlier commitment. Measure how many decisions require clarification, how many deferred records lack an owner, and whether any response implies completion before confirmation. Revise those specific failures before relying on the fallback during an outage.

Integration reliabilityDegraded modesOperational resilienceIncident recovery

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