How to read integration failure rates without drowning in noise
Failure rates look simple until you plot them against traffic volume, partner maintenance windows, and retry storms. A spike at 02:00 may be a batch job colliding with a rate limit—not a broken connector.
Start with a baseline window of four weeks of normal operation. Then score each integration on three axes: hard failures (data never arrived), soft failures (arrived late or incomplete), and silent skips (expected events that never fired).
Weight by business impact. A checkout sync that fails once an hour matters more than a marketing enrichment job that fails once a day. Health reporting should reflect that priority, not treat every endpoint as equal.
Finally, report trends as direction, not drama. A rising soft-failure slope over three weeks is a clearer signal than a single red day on a dashboard.