Canary rollouts, SLO-based analysis, Kubernetes operator patterns, and SRE lessons from the field.
Error-rate alerts during canary analysis cause both false positives and false negatives. Budget burn rate tells you what actually matters: how fast you're spending your reliability capital.
Argo Rollouts is a mature tool with strong GitOps support. Kubestead makes different tradeoffs around opinionated analysis. Here's when you'd pick each.
Hard error thresholds (p99 > 200ms = fail) are brittle. SLO budgets give you a dynamic gate that adjusts to your baseline. Here's how to model it in a RolloutPolicy.
Blue-green is great for a single atomic swap. Progressive delivery is great for observing real traffic at partial exposure. The choice hinges on your blast-radius tolerance.
Not all Prometheus metrics are equally useful for rollout decisions. We walk through the 5 query patterns that consistently separate "looks fine" from "will fail at scale."
Replicated routing, header-based targeting, shadow traffic, and replica-fraction splitting. Each pattern suits a different failure mode tolerance. We explain when to use each.
