progressive-delivery

Canary releases with automatic promotion and rollback using Flagger on Flux. Shifts traffic through Gateway API HTTPRoute weights (Envoy Gateway for north-south, Linkerd for east-west), gates each ste

SKILL.md

GitHub

Ship a new version to a slice of real traffic first, promote it only on evidence, and roll it back without a human when the evidence says no.

Read references/progressive-delivery.md before responding. It holds how a Flagger analysis runs, the provider settings for Envoy Gateway and Linkerd, metric rules, GitOps ownership, rollback by situation, and troubleshooting. The working manifests are in examples/progressive-delivery.

This skill is interactive by default. With no arguments, ask how traffic reaches the app (shared Envoy Gateway, Linkerd east-west, both), where metrics come from (Linkerd viz, gateway Prometheus, other), and whether the app has an HPA, then run canary. Never apply resources or change a live Canary yourself.

Mode dispatch

Mode What it does
install Flagger via Flux from the signed OCI chart, with meshProvider and metricsServer for the user's traffic and metrics
canary Canary and MetricTemplates for one app, from examples/progressive-delivery, adapted to its ports, hostname, listener and HPA
metrics Design or fix the MetricTemplate queries so they measure the canary pods only, and prove they return data
debug Why a canary is stuck, never starts, or always rolls back, from Flagger events and route status
rollback The right rollback for the situation: automatic, revert in Git, skipAnalysis, or suspend

Mode: canary

  1. Start from examples/progressive-delivery/app/canary.yaml. Set targetRef, service.port and targetPort, hosts, and gatewayRefs (the Gateway and sectionName for north-south; the Service for Linkerd east-west).
  2. Add autoscalerRef when the app has an HPA.
  3. Pick interval, stepWeight, maxWeight and threshold from the app's traffic: a check needs enough requests per interval to be meaningful; a low-traffic app needs a load-test webhook or a longer interval.
  4. If the Canary has webhooks, include examples/progressive-delivery/app/loadtester.yaml in the app namespace; the webhook URL must name that Service.
  5. Remove the app's hand-written HTTPRoute in the same change; say so explicitly.
  6. Validate with kubeconform -strict against the CRD catalog, as the example README shows.

Mode: rollback

Use the reference's rollback table. State which objects Flagger owns (primary, generated Services and HTTPRoute) and that the fix for a failed release is always a change in Git, because Flux will keep applying whatever Git says.

Output format

Every mode ends with:

  • Changes, as exact files and fields
  • Blast radius, in one line (which traffic and for how long)
  • Validation, as commands (kubectl describe canary, the metric query, route status)
  • Rollback, for the situation in question