Reconciliation

Reconciliation is a periodic full comparison between two systems that are supposed to agree, run specifically to find what the event stream lost. It exists because no delivery mechanism is perfect over a long enough period: events get dropped during a deploy, rejected by validation and never retried, processed out of order, or written to a queue nobody drained. Each of those is individually rare and collectively certain, and none of them announces itself. It is not a backfill. A backfill is the one-time load of history at the moment an integration is switched on. Reconciliation is the recurring audit that runs forever afterwards, and the two are usually confused because they read the same data. The distinction that matters operationally is that a backfill has an end and reconciliation is a standing cost you should budget for at the point you design the integration, not after the first incident. The design decisions are few and consequential. Compare on a stable identifier that both systems store, never on a name or an email address, because people edit those and the comparison then reports a difference that is really a rename. Compare a deliberately small set of fields — the ones a wrong value would actually hurt — rather than everything, since a wide comparison generates noise that gets ignored and then the whole sweep gets ignored. Decide explicitly whether the sweep reports or repairs: a repairing sweep that trusts the wrong side can overwrite good data at a scale no human error would reach, so a repair needs a designated source of truth per field and a limit on how much it may change in one run. Whatever it does, have it record every difference it finds, including the ones it fixes. The count of discrepancies per run is one of the few honest health signals a pipeline produces — a stream that has quietly degraded shows up as a rising number long before anyone reports missing data, and a sweep that suddenly finds nothing is more often broken than it is a success.