k8s-cilium-troubleshooting

Use when networking fails on a Kubernetes cluster that runs Cilium as its CNI: traffic between pods dropped or timing out after a CiliumNetworkPolicy or CiliumClusterwideNetworkPolicy, "Policy denied"

SKILL.md

GitHub

Cilium troubleshooting

Find out why traffic fails on a cluster that runs Cilium, and propose a fix. Read first, ask before changing. Run the investigation commands freely. Run a command under an "Ask first" line only after the user approves that exact command; if your credentials do not allow it, give it to the user to run. Never run a "Proposed fix"; hand it to the user.

Cilium gives each pod a CiliumEndpoint with a security identity derived from its labels, and enforces Kubernetes NetworkPolicies, CiliumNetworkPolicies (one namespace) and CiliumClusterwideNetworkPolicies by identity. Once any policy selects an endpoint for a direction (ingress or egress), only what some policy allows gets through in that direction. L7 rules (HTTP, DNS) send the traffic through Cilium's Envoy proxy. Hubble records each flow with its verdict.

The commands use placeholders: my-namespace, my-app (the server) and my-web (the client), my-policy, and cilium-xxxxx (the Cilium agent pod on the node of the pod you investigate). Replace them with the real names. The built-in view role cannot read cilium.io objects (Cilium adds nothing to it); if they are forbidden, say so and ask the user for them. Never read Secrets (Cilium keeps its CA and Cluster Mesh credentials in Secrets).

When to use

  • Pods cannot reach each other, or a Service, on a cluster with Cilium; often right after a policy change.
  • DNS lookups fail (bad address, time-outs) after an egress policy.
  • Requests get HTTP 403 Access denied although the app does not return that.
  • Pods stay ContainerCreating with a Cilium CNI error, or Cilium's own pods are not ready.

Hand-offs, by skill name. When the evidence points elsewhere, continue with that skill; if it is not installed, keep going with the steps here and tell the user which skill would help:

  • The Service has no endpoints, a wrong selector or port, or the target pods are not Ready: k8s-network-troubleshooting (Services and Kubernetes NetworkPolicies) and k8s-workload-troubleshooting.
  • A node is NotReady: k8s-node-troubleshooting.
  • Cilium's Gateway API support (the class io.cilium/gateway-controller) does not accept a Gateway or a route: k8s-gateway-api-troubleshooting.

Investigate

Work through the steps in order. Stop when you have the cause and its evidence.

1. Check Cilium itself: the agent on each node, the operator, Envoy.

kubectl get pods -n kube-system -l k8s-app=cilium -o wide
kubectl get deployments,daemonsets -n kube-system | grep -i -E "cilium|hubble"
kubectl get ciliumnodes

2. Read the client's error and the endpoints. A time-out usually means a drop by policy; bad address means DNS; a 403 with Access denied comes from an L7 rule.

kubectl logs -n my-namespace deployment/my-web --tail=20
kubectl get ciliumendpoints -n my-namespace
kubectl get pods -n my-namespace --show-labels -o wide

A CiliumEndpoint should be ready. Pods without one (hostNetwork pods, pods started before Cilium) are not managed and get no policy.

3. Find the policies that select the pods, in both directions: ingress on the server, egress on the client, and cluster-wide ones.

kubectl get ciliumnetworkpolicies,networkpolicies -n my-namespace
kubectl get ciliumclusterwidenetworkpolicies
kubectl get ciliumnetworkpolicy my-policy -n my-namespace -o yaml
kubectl describe ciliumnetworkpolicy my-policy -n my-namespace

Compare each policy's endpointSelector with the pods' labels, and its fromEndpoints, toEndpoints, toPorts and rules with the traffic that fails.

4. Ask Cilium what it drops. Find the agent pod on the node of the client or server (step 1 and the pod's node), then read the recent dropped flows and the endpoints' policy state. These run a reading command inside the agent, so ask first:

Ask first: the agent runs this only after the user approves it.

kubectl exec cilium-xxxxx -n kube-system -c cilium-agent -- hubble observe --last 50 --namespace my-namespace --verdict DROPPED
kubectl exec cilium-xxxxx -n kube-system -c cilium-agent -- cilium-dbg endpoint list

hubble observe needs Hubble, which is on by default. Without it, cilium-dbg status and the agent's logs are the fallback.

5. Open the matching reference and follow its steps:

What you see Reference
Time-outs, Policy denied drops, a policy that allows the wrong peer, default deny surprises, DNS blocked by an egress policy, L7 403s, toFQDNs references/policies.md
Cilium pods not ready, unable to connect to Cilium agent, endpoints not ready, IPAM exhausted, kube-proxy replacement, MTU references/agent-and-datapath.md

Read the results

What the common findings mean (Cilium 1.20):

  • Client time-outs, and Hubble shows Policy denied DROPPED with INGRESS DENIED from the client to the server: a policy selects the server for ingress and none allows this client. Compare the policy's fromEndpoints labels with the client's labels.
  • The same with EGRESS DENIED: a policy selects the client for egress and does not allow the destination.
  • bad address or DNS time-outs right after an egress policy on the client: the policy put the client in default-deny for egress and does not allow DNS to kube-dns (UDP and TCP 53, k8s-app: kube-dns in kube-system).
  • HTTP 403 with the body Access denied: an L7 HTTP rule allows the connection but not that method or path; Cilium's Envoy answers instead of the app.
  • cilium-dbg endpoint list shows Enabled for a direction: some policy selects that endpoint there, so default deny applies to it. Disabled means no policy selects it in that direction.
  • A CiliumNetworkPolicy with VALID False: Cilium rejected it; describe shows why.

State the cause in one sentence, then the evidence you saw (the policy and its selectors, the pods' labels, the drop verdict, the client's error), then the proposed fix.

Proposed fix

Show each fix to the user before running it: what it changes and how to undo it. Creating, changing or deleting a network policy changes what traffic is allowed, often for more pods than the one that fails: never run it yourself; hand the change to the user as a proposed fix, with the narrowest rule that allows the traffic. Changing Cilium's configuration (cilium-config, Helm values) and restarting Cilium are for the cluster's administrators. Prefer fixing the policy in the source it is deployed from. Write each undo with the values you found.

Allow my-web into my-app on port 80, next to the existing rule (the user applies it):

Proposed fix: review before running. The agent does not run this.

kubectl apply -f my-app-ingress-from-my-web.yaml
# Undo: kubectl delete -f my-app-ingress-from-my-web.yaml

with my-app-ingress-from-my-web.yaml a CiliumNetworkPolicy in my-namespace whose endpointSelector is app: my-app and whose ingress allows fromEndpoints app: my-web on toPorts port 80/TCP. Policies add up, so the existing rule stays.

Sources

Checked against the Cilium 1.20 documentation and source (v1.20.4):