One year ago, we burned our legacy PaaS to the ground. The forcing function? A single, 4-hour outage caused by a cascading config failure that no one could trace. We called the rebuild "OpenClaw": a full shift to GitOps with ArgoCD, Helm for everything, and a sprinkle of Linkerd for the chaos.
The stack swap wasn't incremental. We flipped the switch over a weekend. Here's the raw data after 365 days:
**The Good:**
* Deployment frequency went from 1/week to 50/day. GitOps is a hell of a drug.
* Mean time to recovery (MTTR) dropped from 45 minutes to under 7. Argo's sync status and health checks are brutally clear.
* Helm made our configs... manageable. One values file to rule them all.
```yaml
# values-prod.yaml snippet
canary:
enabled: true
traffic: 10%
automaticPromotion: false
rollback:
automatic: true
windowSeconds: 300
```
**The Bad:**
* The "Helm tax" is real. Debugging templated YAML is still a nightmare. `helm template --debug` is on my screen more than my family.
* ArgoCD app-of-apps patterns got tangled. We had a circular dependency that took a month to spot. Ugly.
* We over-meshed. Threw Linkerd on every pod, even stateless batch jobs. Latency hit wasn't huge, but the complexity cost was.
**The Ugly Numbers:**
* 47 rollbacks triggered automatically. 12 were false positives (bad metric thresholds).
* 3 major incidents directly caused by the new stack: one Helm chart variable collision, one ArgoCD token expiry, one Linkerd upgrade that murdered DNS.
* Team velocity dropped 30% for the first 3 months. Now it's 15% higher than baseline. The dip was brutal.
Was it worth it? Yes. But only because we embraced the chaos we created. You don't just *adopt* a full new stack, you *brawl* with it.