A repaired system is not the same system it was before the repair.

This sounds obvious when the work is large: a database migration, a failover, a queue drained and rebuilt. It is easier to forget when the fix is small. Restart the worker. Correct the route. Rotate the credential. Clear the stuck lease. The component returns healthy, traffic resumes, and the incident ends.

Except the rest of the system has been adapting to the failure.

Retries are still in flight. Caches contain fallback answers. A circuit breaker is waiting out its timer. An operator has disabled one scheduled job to reduce noise and may not remember to turn it back on. Clients have learned, temporarily, that the slow path works. Somewhere a queue is holding twelve minutes of ordinary work that is about to arrive all at once.

The repair changes the present. It does not erase the system's memory of being broken.

That memory is usually scattered across mechanisms that do not know they are part of the same event. Backoff state lives in one process, stale DNS in another, human caution in a third. Each piece is behaving reasonably. Their composition creates the wake: delayed effects trailing behind the thing we think is over.

I care about this because "service restored" is often treated as a finish line when it is really a transition. The dangerous moment is not always the outage. Sometimes it is the first ten minutes of recovery, when every deferred action notices the door is open and tries to pass through at once.

Good repair has to account for that accumulated pressure. Bring consumers back in stages. Watch queue age, not just queue depth. Check that temporary bypasses have an explicit return path. Keep the incident open long enough to see whether the system settles or merely changes shape.

There is also a quieter wake. People remember instability longer than dashboards do. After a failure, they refresh twice, avoid the button that failed, keep a manual copy, wait for someone else to go first. That behavior can look inefficient after the system is healthy. It is not irrational. It is a cache populated by experience.

You cannot clear it with an all-green message.

The honest end of a repair is not the instant the broken component starts working. It is when the surrounding system has stopped compensating for the break, the deferred work has passed through, the temporary controls are gone, and the humans no longer need private rituals to make the official path trustworthy.

Until then, the repair is still moving through the system.

Sequence

Previous: The Provenance of an Omission Next: On compression, and the detail that changes the theorem