A partial integration failure needs reconciliation before repair. Determine which effects actually occurred, which are missing, and which remain uncertain so recovery does not duplicate records or overwrite later staff work.
Reconciliation compares expected and actual state
Record reconciliation checks the intended business operation against authoritative records across involved systems. It is not simply rerunning the original workflow. Completed, failed, stale, and unknown steps need different treatment.
For a proposed reconciler, use durable operation identities, relevant source versions, target references, and field-ownership rules. Read the current state and classify differences before proposing a write. Keep manual review for relationships that cannot be established confidently.
Do not assume the source always wins. Staff may legitimately have amended the target after the failed operation. A stale replay can erase those newer decisions unless the conflict policy protects them.
Follow a fictional request sync failure
The source creates a service request and the target creates its matching record. Linking the source reference back to the target fails. A retry that creates a new target record would duplicate the request even though the missing step is only the relationship update.
Reconciliation locates the existing target through the supported operation reference, verifies the match, and repairs the missing link. If several candidates exist, the case enters an authorised review queue rather than choosing by name similarity alone.
| Difference found | Recovery action |
|---|---|
| Target absent, operation not executed | Eligible creation under current rules |
| Target exists, link absent | Repair verified relationship |
| Target has newer staff edit | Apply conflict policy, not blind replacement |
| External outcome unknown | Seek authoritative evidence |
| Duplicate target records | Controlled merge or staff review |
A replayable fixture should include a failure after each boundary and a later staff edit. Assert both the intended final record and preservation of unrelated newer information.
Make repairs auditable and bounded
Record the discrepancy, evidence, proposed change, actor, and final result. Use duplicate protection for repair operations too. A reconciliation job can fail after committing its own update and create the same uncertainty as the original workflow.
Broad full-record comparison can be expensive and noisy. Focus first on authoritative relationships and fields material to the business outcome. Schedule wider checks according to risk and system limits, with an owner for unresolved differences.
This reconciliation design is proposed. The voice integration case study describes related integration work. Pair it with partial-success handling.
Partial integration FAQ
Should the whole workflow be replayed after one failure?
Only if its documented operation semantics make that safe. Usually reconcile completed steps first.
Can email or name matching establish the target record?
Use it only through the approved matching process and retain uncertainty where evidence is insufficient.
What should close a discrepancy?
Verified correct state or a documented disposition with an accountable owner for any remaining limitation.