rollout:
strategy: canary
steps: [5, 25, 50, 100]
dwell: 4m
gate:
p95_latency: < 200ms
error_rate: < 0.5%
Dwell time
Dwell time should be long enough for the gate signals to be statistically meaningful at that traffic share. On low-traffic services, increase it.
Rollback
The previous release is kept warm for the duration of the rollout. Rollback shifts traffic back to it rather than re-running the pipeline.