Deploys are cheap, judgment is not
Most teams ship a canary, open a dashboard and decide by eye. That depends on who is watching, how tired they are and which graph they happened to open.
GitHub Action · canary analysis
deploy-gate compares a canary's real metrics with the version already serving traffic and returns promote rollback or extend, with a report that says why. Missing or thin data never promotes.
Most teams ship a canary, open a dashboard and decide by eye. That depends on who is watching, how tired they are and which graph they happened to open.
"Alert if errors exceed 1%" fires at peak traffic and stays quiet at night. Comparing the canary with the current version over the same minutes cancels those patterns out.
With little traffic, one bad request looks like a 5% error rate. A check fails only when the canary is worse by more than the tolerance you chose, with confidence, so ordinary noise never rolls anything back.
"Ready" is a gate the system passes, not a feeling.
Select a step to see what happens there.
Synthetic traffic, real decision code: this page runs the same src/core the
action ships. Pick a scenario or move the sliders.
| Check | Verdict | p-value | Reason |
|---|
baseline p99 canary p99 (per minute)
Create an IAM role GitHub can assume through OIDC, allowed only
cloudwatch:GetMetricData. No long-lived keys; the action never writes to
AWS.
# .github/deploy-gate.yml
source: cloudwatch
maxDurationMinutes: 30
checks:
- { name: errors, kind: error-rate, maxAbsoluteIncrease: 0.005 }
- { name: latency-p99, kind: latency, statistic: p99, maxRelativeIncrease: 0.10 }
groups:
canary: { dimensions: { TargetGroup: targetgroup/app-canary/abc, LoadBalancer: app/main/123 } }
baseline: { dimensions: { TargetGroup: targetgroup/app-stable/def, LoadBalancer: app/main/123 } }
- id: gate
uses: pablomagnaghi/deploy-gate@v1
with:
canary-started-at: ${{ steps.canary.outputs.started-at }}
fail-on-rollback: false
- if: steps.gate.outputs.decision == 'promote'
run: ./scripts/shift-traffic.sh 100
- if: steps.gate.outputs.decision == 'rollback'
run: ./scripts/shift-traffic.sh 0
On extend, wait and call it again: the window grows
until the checks have enough data or the time limit rolls back. The
examples/ folder wires this up for ALB weights, CodeDeploy and Lambda
aliases.
Every run writes a job summary: the decision, the reason, and one row per check with the numbers and the p-value, so anyone can verify it by hand.