Skip to main content

Cleanup and troubleshooting

1. Cleanup​

1.1 Delete the gateway config​

Deleting the InferenceGatewayConfig garbage collects the whole data plane it created — the Gateway, the Envoy proxy and service, the BBR deployment, the endpoint pickers, and the generated InferencePool and HTTPRoute resources:

kubectl delete inferencegatewayconfig "${GATEWAY_NAME}" -n "${MODEL_NS}"

Confirm the data plane is gone:

kubectl get pods -n "${MODEL_NS}" | grep -E 'epp|bbr'
kubectl get svc -n "${SYSTEM_NS}" | grep envoy

Both should return nothing. Your model deployments keep running — the gateway and the models have independent lifecycles.

1.2 Delete the model backends​

kubectl delete inferenceendpointconfig vllm-qwen -n "${MODEL_NS}"
kubectl delete deploy vllm-qwen-small -n "${MODEL_NS}"

1.3 Delete the namespace​

kubectl delete namespace "${MODEL_NS}"
info

The add-on itself is left in place. Removing it would tear down the control plane shared by every gateway in the cluster.

2. Troubleshooting​

2.1 Gateway config rejected on apply​

bbr must be enabled when more than one scheduler is defined

Set bbr.enabled: true. This is required for any config with two or more schedulers, and is enforced at admission.

strict decoding error: unknown field "spec.<name>"

The CRD schema is closed, so unknown fields are rejected rather than ignored. Check the field against the live schema:

kubectl explain inferencegatewayconfig.spec --recursive

Validate a config without persisting it:

kubectl apply --dry-run=server -f your-config.yaml

2.2 Requests return 404​

no matching model: set the model in the request body or the x-gateway-model-name header

The gateway found no scheduler whose modelName matches the request body's model value. Compare the two:

kubectl get inferencegatewayconfig "${GATEWAY_NAME}" -n "${MODEL_NS}" \
  -o jsonpath='{range .spec.schedulers[*]}{.name}{" -> "}{.modelName}{"\n"}{end}'

A vLLM NotFoundError instead of the gateway message

The request reached a backend, which then rejected the model name. With a single scheduler and BBR disabled the gateway does no model matching and forwards everything to its one pool, so a bad model name surfaces at the backend rather than at the gateway.

2.3 Gateway ready but requests fail​

Check that the scheduler's modelSelector actually matches running pods:

kubectl get pods -n "${MODEL_NS}" -l app=vllm-qwen
kubectl get inferencepool -n "${MODEL_NS}"
kubectl get endpointslice -n "${MODEL_NS}"

A scheduler whose selector matches nothing still reports Ready, because the config is valid — there are simply no endpoints to route to.

2.4 First request fails right after deploying a model​

The pod is Running but vLLM is still loading. This happens when a hand-written Deployment has no readiness probe, in which case kubectl rollout status returns success prematurely. Add one:

readinessProbe:
  httpGet:
    path: /ping
    port: 8000
  periodSeconds: 10
  failureThreshold: 3

Models deployed through the inference operator get an equivalent probe automatically.

2.5 Cannot resolve the gateway hostname​

curl: (6) Could not resolve host: envoy-...svc.cluster.local

The .svc.cluster.local name resolves only inside the cluster. Run curl from a pod, as every example here does. The load balancers the gateway provisions are internal in scheme, so they are reachable from inside the VPC but not from a laptop.

2.6 Cannot find POST lines in backend logs​

Model pods emit roughly 40 GET /metrics lines per second, and kubectl logs -l <selector> defaults to --tail=10. Request all lines and identify the pod:

kubectl logs -n "${MODEL_NS}" -l app=vllm-qwen --tail=-1 --prefix \
  | grep 'POST /v1/completions'

3. Useful commands​

# Gateway status conditions
kubectl get inferencegatewayconfig "${GATEWAY_NAME}" -n "${MODEL_NS}" \
  -o jsonpath='{range .status.conditions[*]}{.type}={.status}{"\n"}{end}'

# Controller reconciliation log
kubectl logs -n "${SYSTEM_NS}" deploy/inference-gateway-controller --tail=200

# Routing evidence from an endpoint picker
# --since is a lookback window; widen it if the request was issued a while ago
kubectl logs -n "${MODEL_NS}" -l app=qwen-epp --tail=-1 --since=15m \
  | grep -oE '"x-request-id":"[a-f0-9-]+","modelName":"[^"]+"'

# Per-pod request and prefix-cache counters
curl -s http://<pod-ip>:8000/metrics \
  | grep -E '^vllm:(request_success_total|gpu_prefix_cache_hits_total|gpu_prefix_cache_queries_total)'