2 June 2026

Retries that look like reliability

How silent backoff can inflate success rates while pushing work past your SLA window.

Retries are kindness to flaky networks. They are also a favourite place for latency to hide.

When every attempt is collapsed into a single “success” event, you lose the story of how many times the automation asked and how long it waited between asks. A payment sync that succeeds on the fourth try after forty minutes is not the same animal as a sync that succeeds on the first try in four seconds.

A practical minimum: count attempts, store the last upstream status, and emit a duration that includes backoff. Then set an alert on attempt volume, not only on final failure. You will catch partner degradation while there is still time to pause non-critical lanes.