Although Kyverno’s goal is to make policy simple, sometimes trouble still strikes. The following sections can be used to help troubleshoot and recover when things go wrong.
API server is blocked
Symptom: Kyverno Pods are not running and the API server is timing out due to webhook timeouts. My cluster appears “broken”.
Cause: This can happen if all Kyverno Pods are down, due typically to a cluster outage or improper scaling/killing of full node groups, and policies were configure to fail-closed while matching on Pods. This is usually only the case when the Kyverno Namespace has not been excluded (not the default behavior) or potentially system Namespaces which have cluster-critical components such as
Solution: Delete the Kyverno validating and mutating webhook configurations. When Kyverno recovers, check your Namespace exclusions. Follow the steps below. Also consider running the admission controller component with 3 replicas.
- Delete the validating and mutating webhook configurations that instruct the API server to forward requests to Kyverno:
1kubectl delete validatingwebhookconfiguration kyverno-resource-validating-webhook-cfg 2kubectl delete mutatingwebhookconfiguration kyverno-resource-mutating-webhook-cfg
Note that these two webhook configurations are used for resources. Other Kyverno webhooks are for internal operations and typically do not need to be deleted. When Kyverno recovers, its webhooks will be recreated based on the currently-installed policies.
- Restart Kyverno
This step is typically not necessary. In case it is, either delete the Kyverno Pods or scale the Deployment down to zero and then up. For example, for an installation with three replicas in the default Namespace use:
1kubectl scale deploy kyverno-admission-controller -n kyverno --replicas 0 2kubectl scale deploy kyverno-admission-controller -n kyverno --replicas 3
- Consider excluding namespaces
Use Namespace selectors to filter requests to system Namespaces. Note that this configuration bypasses all policy checks on select Namespaces and may violate security best practices. When excluding Namespaces, it is your responsibility to ensure other controls such as Kubernetes RBAC are configured since Kyverno cannot apply any policies to objects therein. For more information, see the Security vs Operability section. The Kyverno Namespace is excluded by default. And if running Kyverno on certain PaaS platforms, additional Namespaces may need to be excluded as well, for example
Policies are not applied
Symptom: My policies are created but nothing seems to happen when I create a resource that should trigger them.
Solution: There are a few moving parts that need to be checked to ensure Kyverno is receiving information from Kubernetes and is in good health.
Check and ensure the Kyverno Pod(s) are running. Assuming Kyverno was installed into the default Namespace of
kyverno, use the command
kubectl -n kyverno get poto check their status. The status should be
Runningat all times.
Check all the policies installed in the cluster to ensure they are all reporting
1$ kubectl get cpol,pol -A 2NAME BACKGROUND VALIDATE ACTION READY AGE MESSAGE 3inject-entrypoint true Audit True 15s Ready
Kyverno registers as two types of webhooks with Kubernetes. Check the status of registered webhooks to ensure Kyverno is among them.
1$ kubectl get validatingwebhookconfigurations,mutatingwebhookconfigurations 2 NAME WEBHOOKS AGE 3 validatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-cleanup-validating-webhook-cfg 1 5d21h 4 validatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-policy-validating-webhook-cfg 1 5d21h 5 validatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-exception-validating-webhook-cfg 1 5d21h 6 validatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-resource-validating-webhook-cfg 1 5d21h 7 8 NAME WEBHOOKS AGE 9 mutatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-policy-mutating-webhook-cfg 1 5d21h 10 mutatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-verify-mutating-webhook-cfg 1 5d21h 11 mutatingwebhookconfiguration.admissionregistration.k8s.io/kyverno-resource-mutating-webhook-cfg 1 5d21h
The age should be consistent with the age of the currently running Kyverno Pod(s). If the age of these webhooks shows, for example, a few seconds old, Kyverno may be having trouble registering with Kubernetes.
Test that name resolution and connectivity to the Kyverno service works inside your cluster by starting a simple
busyboxPod and trying to connect to Kyverno. Enter the
wgetcommand as shown below. If the response is not “remote file exists” then there is a network connectivity or DNS issue within your cluster. If your cluster was provisioned with kubespray, see if this comment helps you.
1$ kubectl run busybox --rm -ti --image=busybox -- /bin/sh 2If you don't see a command prompt, try pressing enter. 3/ # wget --no-check-certificate --spider --timeout=1 https://kyverno-svc.kyverno.svc:443/health/liveness 4Connecting to kyverno-svc.kyverno.svc:443 (100.67.141.176:443) 5remote file exists 6/ # exit 7Session ended, resume using 'kubectl attach busybox -c busybox -i -t' command when the pod is running 8pod "busybox" deleted
validatepolicies, ensure that
validationFailureActionis set to
Enforceif your expectation is that applicable resources should be blocked. Most policies in the samples library are purposefully set to
Auditmode so they don’t have any unintended consequences for new users. It could be that, if the prior steps check out, Kyverno is working fine only that your policy is configured to not immediately block resources.
Check and ensure you aren’t creating a resource that is either excluded from Kyverno’s processing by default, or that it hasn’t been created in an excluded Namespace. Kyverno uses a ConfigMap by default called
kyvernoin the Kyverno Namespace to filter out some of these things. The key name is
resourceFiltersand more details can be found here.
Check the same ConfigMap and ensure that the user/principal or group responsible for submission of your resource is not being excluded. Check the
excludeUsernamesand others if they exist.
Kyverno consumes a lot of resources or I see OOMKills
Symptom: Kyverno is using too much memory or CPU. How can I understand what is causing this?
Solution: It is important to understand how Kyverno experiences and processes work to know if what you deem as “too much” is, in fact, too much. Kyverno dynamically configures its webhooks (by default but configurable) according the policies which are loaded and on what resources they match. There is no easy rubric to follow where resource requirements are directly proportional to, for example, number of Pods or Nodes in a cluster. The following questions need to be asked and answered to build a full picture of the resources consumed by Kyverno.
- What policies are in the cluster and on what types of resources do they match? Policies which match on wildcards (
"*") cause a tremendous load on Kyverno and should be avoided if possible as they instruct the Kubernetes API server to send to Kyverno every action on every resource in the cluster. Even if Kyverno does not have matching policies for most of these resources, it is required to respond to every single one. If even one policy matches on a wildcard, expect the resources needed by Kyverno to easily double, triple, or more.
- Which controller is experiencing the load? Each Kyverno controller has different responsibilities. See the controller guide for more details. Each controller can be independently scaled, but before immediately scaling in any direction take the time to study the load.
- Are the default requests and limits still in effect? It is possible the amount of load Kyverno (any of its controllers) is experiencing is beyond the capabilities of the default requests and limits. These defaults have been selected based on a good mix of real-world usage and feedback but may not suit everyone. In extremely large and active clusters, from Kyverno’s perspective, you may need to increase these.
- What do your monitoring metrics say? Kyverno is a critical piece of cluster infrastructure and must be monitored effectively just like other pieces. There are several metrics which give a sense of how active Kyverno is, the most important being admission request count. Others include consumed memory and CPU utilization. Sizing should always be done based on peak consumption and not averages.
You can also follow the steps on the Kyverno wiki for enabling memory and CPU profiling.
Symptom: I’m using AKS and Kyverno is using too much memory or CPU or produces many audit logs
Solution: On AKS the Kyverno webhooks will be mutated by the AKS Admissions Enforcer plugin, that can lead to an endless update loop. To prevent that behavior, set the annotation
"admissions.enforcer/disabled": true to all Kyverno webhooks. When installing via Helm, the annotation can be added with
Kyverno is slow to respond
Symptom: Kyverno’s operation seems slow in either mutating resources or validating them, causing additional time to create resources in the Kubernetes cluster.
Solution: Check the Kyverno logs for messages about throttling. If many are found, this indicates Kyverno is making too many API calls in too rapid a succession which the Kubernetes API server will throttle. Increase the values, or set the flags,
--clientRateLimitBurst. While these flags have very sensible values after much field trials, in some cases they may need to be increased.
Policies are partially applied
Symptom: Kyverno is working for some policies but not others. How can I see what’s going on?
Solution: The first thing is to check the logs from the Kyverno Pod to see if it describes why a policy or rule isn’t working.
Check the Pod logs from Kyverno. Assuming Kyverno was installed into the default Namespace called
kyvernouse the command
kubectl -n kyverno logs <kyverno_pod_name>to show the logs. To watch the logs live, add the
-fswitch for the “follow” option.
If no helpful information is being displayed at the default logging level, increase the level of verbosity by editing the Kyverno Deployment. To edit the Deployment, assuming Kyverno was installed into the default Namespace, use the command
kubectl -n kyverno edit deploy kyverno-<controller_type>-controller. Find the
argssection for the container named
kyvernoand either add the
-vswitch or increase to a higher level. The flag
-v=6will increase the logging level to its highest. Take care to revert this change once troubleshooting steps are concluded.
Symptom: I have a large cluster with many objects and many Kyverno policies. Kyverno is seen to sometimes crash.
Solution: In cases of very large scale, it may be required to increase the memory limit of the Kyverno Pod so it can keep track of these objects.
- First, see the above troubleshooting section. If changes are required, edit the necessary Kyverno Deployment and increase the memory limit on the container. Change the
resources.limits.memoryfield to a larger value. Continue to monitor the memory usage by using something like the Kubernetes metrics-server.
Kyverno fails on GKE
Symptom: I’m using GKE and after installing Kyverno, my cluster is either broken or I’m seeing timeouts and other issues.
Solution: Private GKE clusters do not allow certain communications from the control planes to the workers, which Kyverno requires to receive webhooks from the API server. In order to resolve this issue, create a firewall rule which allows the control plane to speak to workers on the Kyverno TCP port which, by default at this time, is 9443.
Kyverno fails on EKS
Symptom: I’m an EKS user and I’m finding that resources that should be blocked by a Kyverno policy are not. My cluster does not use the VPC CNI.
Solution: When using EKS with a custom CNI plug-in (ex., Calico), the Kyverno webhook cannot be reached by the API server because the control plane nodes, which cannot use a custom CNI, differ from the configuration of the worker nodes, which can. In order to resolve this, when installing Kyverno via Helm, set the
hostNetwork option to
true. See also this note. AWS lists the alternate compatible CNI plug-ins here.
Symptom: When creating Pods or other resources, I receive similar errors like
Error from server (InternalError): Internal error occurred: failed calling webhook "validate.kyverno.svc-fail": Post "https://kyverno-svc.kyverno.svc:443/validate?timeout=10s": context deadline exceeded.
Solution: When using EKS with the VPC CNI, problems may arise if the CNI plug-in is outdated. Upgrade the VPC CNI plug-in to a version supported and compatible with the Kubernetes version running in the EKS cluster.
Symptom: Kyverno pods emit logs stating
Waited for <n>s due to client-side throttling; the creation of mutated resources may be delayed.
Solution: Try increasing
clientRateLimitQPS (documented here). If that doesn’t resolve the problem, you can experiment with slowly increasing these values. Just bear in mind that higher values place more pressure on the Kubernetes API (the client-side throttling was implemented for a reason), which could result in cluster-wide latency, so proceed with caution.
Policy definition not working
Symptom: My policy seems like it should work based on how I have authored it but it doesn’t.
Solution: There can be many reasons why a policy may fail to work as intended, assuming other policies work. One of the most common reasons is that the API server is sending different contents than what you have accounted for in your policy. To see the full contents of the AdmissionReview request the Kubernetes API server sends to Kyverno, add the
dumpPayload container flag set to
true and check the logs. This has performance impact so it should be removed or set back to
false when complete.
The second most common reason policies may fail to operate per design is due to variables. To see the values Kyverno is substituting for variables, increase logging to level
4 by setting the container flag
-v=4. You can
grep for the string
variable (or use tools such as stern) and only see the values being substituted for those variables.
Admission reports are stacking up
Symptom: Admission reports keep accumulating in the cluster, taking more and more etcd space and slowing down requests.
Diagnose: Please follow the troubleshooting docs to determine if you are affected by this issue.
Solution: Admission reports can accumulate if the reports controller is not working properly so the first thing to check is if the reports controller is running and does not continuously restarts. If the controller works as expected, another potential cause is that it fails to aggregate admission reports fast enough. This usually happens when the controller is throttled. You can fix this by increasing QPS and burst rates for the controller by setting
Note that starting with Kyverno 1.10, two cron jobs are responsible for deleting admission reports automatically if they accumulate over a certain threshold.