The situation
Every release was a moment of risk. When a change misbehaved, users felt it before anyone could react, and getting back to a good state took manual work under pressure.
With high-traffic applications on EKS and GKE, that risk grows with every service and every deploy.
What I built
- Managed EKS and GKE clusters and packaged services with Helm charts.
- Integrated ArgoCD and FluxCD so the cluster follows what’s in Git, and a rollback is a revert.
- Implemented blue-green and canary deployment strategies, so new versions get a slice of real traffic before everyone sees them.
- Set up autoscaling policies for the high-traffic applications so a rollout never competes with load.
How it flows
- Merge to main
- GitOps sync
- Canary slice
- Check health
- Promote or roll back
Decisions I made and why
GitOps over pushing from CI
When the cluster pulls its desired state from Git, there’s one source of truth. You can see what’s running, and undoing a change is a normal Git revert instead of a special procedure.
Canary for risky services, blue-green for the rest
Canary releases catch problems that only show up under real traffic. Blue-green is simpler when a clean switch is enough. Using both kept the setup no more complex than each service needed.
Let a small slice find the problem
A new version that misbehaves is caught while only a few users can see it, and everyone else stays on the version that works. People investigate afterwards, but users stop being affected quickly.
What changed
Service interruptions during releases dropped 95%, and releases stopped being something the teams had to schedule around.
What I’d do next
Tie canary analysis to the same SLOs the teams already watch, so “healthy enough to promote” means exactly what it means in production.
Stack Kubernetes (EKS, GKE), Helm, ArgoCD, FluxCD
Want to see the idea in action? The deploy playground runs a small simulation of this setup, including a canary that rolls itself back.
Details about State Farm’s systems are kept general on purpose. The approach, the tools and the results are from my work there.