Dynamic Resource Allocation gives Kubernetes a structured model for GPUs and other accelerators. But after a ResourceClaim is scheduled and handed to a kubelet plugin, most drivers go quiet. When a pod is stuck in ContainerCreating, engineers can't tell whether the driver, the scheduler, or the API server is at fault, and they have no baseline for how long device preparation should take.

This talk walks through the observability layer merged into the upstream dra-example-driver. We'll cover where to instrument (callback boundaries, not the gRPC server), which signals matter (prepare and unprepare counters, latency histograms, fatal background errors), and how to get API client metrics almost for free from component-base. Then we'll use those metrics to trace a failing claim end to end, go over the deployment gotchas, and be clear about what is still missing. You'll leave with a pattern you can reuse for any DRA driver.