Pods scheduled on a Fargate profile behave differently from pods on a managed node group, and most teams carry their old assumptions across without adjusting them. There is no shared node, no DaemonSet, and no host to install anything on. Whatever Kubernetes was doing to protect your workload, it is doing less of it here.

The application-level risk is unchanged and unmeasured. A pod that stalls waiting on a database stays Ready and keeps receiving traffic. Your dashboards report a healthy deployment while requests queue behind a call that will never return, and the first real signal reaches you through a customer complaint.

Gremlin gives you a way to see this before it happens. Add the Failure Flags SDK to your application and deploy the sidecar into the pod, and you can inject latency and errors into specific code paths on a live deployment. You find out whether readiness probes reflect actual health, whether a slow dependency stays contained, and whether the resilience you assumed the platform provides is resilience you have.