github-release-management

GitHub Release Management Skill

Ships releases via progressive canary traffic gates (5% → 25% → 50% → 100%) with auto-rollback on health-gate failures. Implemented as a GitHub Actions workflow that pauses between stages to evaluate error rate, p95 latency, and health-check telemetry before widening the blast radius.

Trigger

Use this skill when:

• Cutting a production release that must not flip 100% of traffic at once
• Introducing a breaking runtime change behind a feature flag
• Wiring GitHub Actions to promote a build through canary gates with auto-rollback
• Debugging a release that stalled mid-canary or rolled back unexpectedly

Canary Traffic Routing

Traffic is shifted by updating the weight on the canary target. Each stage is held long enough to accumulate meaningful telemetry (default: 10 min minimum, 30 min for stage 1).

Stage	Canary Weight	Stable Weight	Min Hold	Gate Required
1	5%	95%	30 min	Yes
2	25%	75%	15 min	Yes
3	50%	50%	15 min	Yes
4	100%	0%	sticky	Final check

Example GitHub Actions step (weight update via router API):

- name: Shift traffic to canary 5%
  run: |
    curl -sSf -X POST "$ROUTER_API/releases/$RELEASE_ID/weights" \
      -H "Authorization: Bearer $ROUTER_TOKEN" \
      -H "Content-Type: application/json" \
      -d '{"canary": 5, "stable": 95}'

Health Gate Criteria

A stage is healthy only when ALL of the following are true over the hold window:

• Error rate: canary 5xx rate <= stable 5xx rate + 0.5 percentage points AND absolute canary 5xx rate < 1.0%
• p95 latency: canary p95 <= stable p95 × 1.20 (max +20% regression tolerated)
• Health checks: 100% of /healthz probes return 200 for the last 5 minutes
• Saturation: canary CPU and memory within 90% of stable baseline