ENOTFOUND, Name or service not known and similar errors mean name resolution failed before the client could connect. In a distributed AI system, the name may identify a model provider, internal inference Service, vector database, tool endpoint or gateway.

Confirm the exact query

Inspect the configured hostname without exposing credentials. Check for a URL passed where a hostname was expected, spelling errors, an empty environment variable or a service name valid only inside another network.

getent hosts MODEL_HOST
dig MODEL_HOST A
dig MODEL_HOST AAAA
dig MODEL_HOST CNAME

Run the checks from the failing container, pod or runtime. Host DNS success does not prove container DNS success.

Separate failure types

• NXDOMAIN: authoritative DNS says the name does not exist;
• SERVFAIL: resolver or authoritative processing failed;
• timeout: resolver is unreachable or responses are blocked;
• wrong address: lookup succeeds but points to stale or unintended infrastructure;
• intermittent failure: caching, resolver saturation, rollout or split-horizon inconsistency.

Do not replace a hostname with a hardcoded IP unless the service contract explicitly requires it. That bypasses rotation, failover and certificate identity.

Kubernetes service discovery

Inside a cluster, verify the Service, namespace and endpoints:

kubectl get svc,endpoints -n AI_NAMESPACE
kubectl exec -n AI_NAMESPACE POD -- cat /etc/resolv.conf
kubectl exec -n AI_NAMESPACE POD -- nslookup MODEL_SERVICE
kubectl get pods -n kube-system -l k8s-app=kube-dns

Check search domains and whether a short service name is being resolved from the expected namespace. If CoreDNS is unhealthy, investigate its Service, endpoints, logs and upstream resolver path rather than restarting application pods repeatedly.

Cloud and private endpoints

Private model endpoints may depend on VPC-specific zones, private resolver rules or VPN connectivity. Confirm the workload runs in the network where that DNS view exists. Review recent endpoint, zone and infrastructure changes.

For public providers, verify current documented API hostnames. Do not construct provider endpoints from model names unless the provider contract says to do so.

Reliability and retries

Retry transient resolver failures with bounded backoff, but treat persistent NXDOMAIN as configuration failure. Cache according to DNS TTL and library behavior; indefinite application caching can preserve stale routing after a deployment.

Monitor failures by hostname, resolver, runtime and deployment version. Add resolver outage and stale-record cases to AI Testing & Evaluation.

Connect DNS ownership to AI Operations, network placement to AI Deployment & Hosting and service contracts to AI Application Architecture. For fundamentals, see DNS architecture for AI services.