Service Mesh Platform · View 18 of 31 · 5 · Runtime
Decisions
- Analysis runs on a 30-second interval and aborts on two consecutive failures. The 41 seconds includes detection; the revert itself is one route change delivered as a delta.
- The rollout reads the canary share from mesh metrics before judging error ratio. If the applied share is outside tolerance, the analysis is inconclusive and the rollout pauses instead of judging a canary that received no traffic.
- A person aborts the Rollout object. Admission guards reject direct edits to a route owned by an active rollout, which is what stops a rollback racing a rollout.
Targets
- Rollback complete ≤ 60 s after breach detection. Weight change applied fleet-wide p95 ≤ 10 s.
Risks
- At 1% of a low-traffic service, an analysis window may hold too few requests to judge. Rollouts declare a minimum request count per step; below it the step waits rather than guesses.
