This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Monitoring, Logging, and Debugging

Set up monitoring and logging to troubleshoot a cluster, or debug a containerized application.

Sometimes things go wrong. This guide is aimed at making them right. It has two sections:

You should also check the known issues for the release you're using.

Getting help

If your problem isn't answered by any of the guides above, there are variety of ways for you to get help from the Kubernetes community.


The documentation on this site has been structured to provide answers to a wide range of questions. Concepts explain the Kubernetes architecture and how each component works, while Setup provides practical instructions for getting started. Tasks show how to accomplish commonly used tasks, and Tutorials are more comprehensive walkthroughs of real-world, industry-specific, or end-to-end development scenarios. The Reference section provides detailed documentation on the Kubernetes API and command-line interfaces (CLIs), such as kubectl.

Help! My question isn't covered! I need help now!

Stack Overflow

Someone else from the community may have already asked a similar question or may be able to help with your problem. The Kubernetes team will also monitor posts tagged Kubernetes. If there aren't any existing questions that help, please ensure that your question is on-topic on Stack Overflow and that you read through the guidance on how to ask a new question, before asking a new one!


Many people from the Kubernetes community hang out on Kubernetes Slack in the #kubernetes-users channel. Slack requires registration; you can request an invitation, and registration is open to everyone). Feel free to come and ask any and all questions. Once registered, access the Kubernetes organisation in Slack via your web browser or via Slack's own dedicated app.

Once you are registered, browse the growing list of channels for various subjects of interest. For example, people new to Kubernetes may also want to join the #kubernetes-novice channel. As another example, developers should join the #kubernetes-dev channel.

There are also many country specific / local language channels. Feel free to join these channels for localized support and info:

Country / language specific Slack channels
China#cn-users, #cn-events
France#fr-users, #fr-events
Germany#de-users, #de-events
India#in-users, #in-events
Italy#it-users, #it-events
Japan#jp-users, #jp-events
Turkey#tr-users, #tr-events


You're welcome to join the official Kubernetes Forum:

Bugs and feature requests

If you have what looks like a bug, or you would like to make a feature request, please use the GitHub issue tracking system.

Before you file an issue, please search existing issues to see if your issue is already covered.

If filing a bug, please include detailed information about how to reproduce the problem, such as:

  • Kubernetes version: kubectl version
  • Cloud provider, OS distro, network configuration, and container runtime version
  • Steps to reproduce the problem

1 - Troubleshooting Applications

Debugging common containerized application issues.

This doc contains a set of resources for fixing issues with containerized applications. It covers things like common issues with Kubernetes resources (like Pods, Services, or StatefulSets), advice on making sense of container termination messages, and ways to debug running containers.

1.1 - Debug Pods

This guide is to help users debug applications that are deployed into Kubernetes and not behaving correctly. This is not a guide for people who want to debug their cluster. For that you should check out this guide.

Diagnosing the problem

The first step in troubleshooting is triage. What is the problem? Is it your Pods, your Replication Controller or your Service?

Debugging Pods

The first step in debugging a Pod is taking a look at it. Check the current state of the Pod and recent events with the following command:

kubectl describe pods ${POD_NAME}

Look at the state of the containers in the pod. Are they all Running? Have there been recent restarts?

Continue debugging depending on the state of the pods.

My pod stays pending

If a Pod is stuck in Pending it means that it can not be scheduled onto a node. Generally this is because there are insufficient resources of one type or another that prevent scheduling. Look at the output of the kubectl describe ... command above. There should be messages from the scheduler about why it can not schedule your pod. Reasons include:

  • You don't have enough resources: You may have exhausted the supply of CPU or Memory in your cluster, in this case you need to delete Pods, adjust resource requests, or add new nodes to your cluster. See Compute Resources document for more information.

  • You are using hostPort: When you bind a Pod to a hostPort there are a limited number of places that pod can be scheduled. In most cases, hostPort is unnecessary, try using a Service object to expose your Pod. If you do require hostPort then you can only schedule as many Pods as there are nodes in your Kubernetes cluster.

My pod stays waiting

If a Pod is stuck in the Waiting state, then it has been scheduled to a worker node, but it can't run on that machine. Again, the information from kubectl describe ... should be informative. The most common cause of Waiting pods is a failure to pull the image. There are three things to check:

  • Make sure that you have the name of the image correct.
  • Have you pushed the image to the registry?
  • Try to manually pull the image to see if the image can be pulled. For example, if you use Docker on your PC, run docker pull <image>.

My pod is crashing or otherwise unhealthy

Once your pod has been scheduled, the methods described in Debug Running Pods are available for debugging.

My pod is running but not doing what I told it to do

If your pod is not behaving as you expected, it may be that there was an error in your pod description (e.g. mypod.yaml file on your local machine), and that the error was silently ignored when you created the pod. Often a section of the pod description is nested incorrectly, or a key name is typed incorrectly, and so the key is ignored. For example, if you misspelled command as commnd then the pod will be created but will not use the command line you intended it to use.

The first thing to do is to delete your pod and try creating it again with the --validate option. For example, run kubectl apply --validate -f mypod.yaml. If you misspelled command as commnd then will give an error like this:

I0805 10:43:25.129850   46757 schema.go:126] unknown field: commnd
I0805 10:43:25.129973   46757 schema.go:129] this may be a false alarm, see

The next thing to check is whether the pod on the apiserver matches the pod you meant to create (e.g. in a yaml file on your local machine). For example, run kubectl get pods/mypod -o yaml > mypod-on-apiserver.yaml and then manually compare the original pod description, mypod.yaml with the one you got back from apiserver, mypod-on-apiserver.yaml. There will typically be some lines on the "apiserver" version that are not on the original version. This is expected. However, if there are lines on the original that are not on the apiserver version, then this may indicate a problem with your pod spec.

Debugging Replication Controllers

Replication controllers are fairly straightforward. They can either create Pods or they can't. If they can't create pods, then please refer to the instructions above to debug your pods.

You can also use kubectl describe rc ${CONTROLLER_NAME} to introspect events related to the replication controller.

Debugging Services

Services provide load balancing across a set of pods. There are several common problems that can make Services not work properly. The following instructions should help debug Service problems.

First, verify that there are endpoints for the service. For every Service object, the apiserver makes an endpoints resource available.

You can view this resource with:

kubectl get endpoints ${SERVICE_NAME}

Make sure that the endpoints match up with the number of pods that you expect to be members of your service. For example, if your Service is for an nginx container with 3 replicas, you would expect to see three different IP addresses in the Service's endpoints.

My service is missing endpoints

If you are missing endpoints, try listing pods using the labels that Service uses. Imagine that you have a Service where the labels are:

  - selector:
     name: nginx
     type: frontend

You can use:

kubectl get pods --selector=name=nginx,type=frontend

to list pods that match this selector. Verify that the list matches the Pods that you expect to provide your Service. Verify that the pod's containerPort matches up with the Service's targetPort

Network traffic is not forwarded

Please see debugging service for more information.

What's next

If none of the above solves your problem, follow the instructions in Debugging Service document to make sure that your Service is running, has Endpoints, and your Pods are actually serving; you have DNS working, iptables rules installed, and kube-proxy does not seem to be misbehaving.

You may also visit troubleshooting document for more information.

1.2 - Debug Services

An issue that comes up rather frequently for new installations of Kubernetes is that a Service is not working properly. You've run your Pods through a Deployment (or other workload controller) and created a Service, but you get no response when you try to access it. This document will hopefully help you to figure out what's going wrong.

Running commands in a Pod

For many steps here you will want to see what a Pod running in the cluster sees. The simplest way to do this is to run an interactive busybox Pod:

kubectl run -it --rm --restart=Never busybox sh

If you already have a running Pod that you prefer to use, you can run a command in it using:

kubectl exec <POD-NAME> -c <CONTAINER-NAME> -- <COMMAND>


For the purposes of this walk-through, let's run some Pods. Since you're probably debugging your own Service you can substitute your own details, or you can follow along and get a second data point.

kubectl create deployment hostnames
deployment.apps/hostnames created

kubectl commands will print the type and name of the resource created or mutated, which can then be used in subsequent commands.

Let's scale the deployment to 3 replicas.

kubectl scale deployment hostnames --replicas=3
deployment.apps/hostnames scaled

Note that this is the same as if you had started the Deployment with the following YAML:

apiVersion: apps/v1
kind: Deployment
    app: hostnames
  name: hostnames
      app: hostnames
  replicas: 3
        app: hostnames
      - name: hostnames

The label "app" is automatically set by kubectl create deployment to the name of the Deployment.

You can confirm your Pods are running:

kubectl get pods -l app=hostnames
NAME                        READY     STATUS    RESTARTS   AGE
hostnames-632524106-bbpiw   1/1       Running   0          2m
hostnames-632524106-ly40y   1/1       Running   0          2m
hostnames-632524106-tlaok   1/1       Running   0          2m

You can also confirm that your Pods are serving. You can get the list of Pod IP addresses and test them directly.

kubectl get pods -l app=hostnames \
    -o go-template='{{range .items}}{{.status.podIP}}{{"\n"}}{{end}}'

The example container used for this walk-through serves its own hostname via HTTP on port 9376, but if you are debugging your own app, you'll want to use whatever port number your Pods are listening on.

From within a pod:

for ep in; do
    wget -qO- $ep

This should produce something like:


If you are not getting the responses you expect at this point, your Pods might not be healthy or might not be listening on the port you think they are. You might find kubectl logs to be useful for seeing what is happening, or perhaps you need to kubectl exec directly into your Pods and debug from there.

Assuming everything has gone to plan so far, you can start to investigate why your Service doesn't work.

Does the Service exist?

The astute reader will have noticed that you did not actually create a Service yet - that is intentional. This is a step that sometimes gets forgotten, and is the first thing to check.

What would happen if you tried to access a non-existent Service? If you have another Pod that consumes this Service by name you would get something like:

wget -O- hostnames
Resolving hostnames (hostnames)... failed: Name or service not known.
wget: unable to resolve host address 'hostnames'

The first thing to check is whether that Service actually exists:

kubectl get svc hostnames
No resources found.
Error from server (NotFound): services "hostnames" not found

Let's create the Service. As before, this is for the walk-through - you can use your own Service's details here.

kubectl expose deployment hostnames --port=80 --target-port=9376
service/hostnames exposed

And read it back:

kubectl get svc hostnames
hostnames   ClusterIP   <none>        80/TCP    5s

Now you know that the Service exists.

As before, this is the same as if you had started the Service with YAML:

apiVersion: v1
kind: Service
    app: hostnames
  name: hostnames
    app: hostnames
  - name: default
    protocol: TCP
    port: 80
    targetPort: 9376

In order to highlight the full range of configuration, the Service you created here uses a different port number than the Pods. For many real-world Services, these values might be the same.

Any Network Policy Ingress rules affecting the target Pods?

If you have deployed any Network Policy Ingress rules which may affect incoming traffic to hostnames-* Pods, these need to be reviewed.

Please refer to Network Policies for more details.

Does the Service work by DNS name?

One of the most common ways that clients consume a Service is through a DNS name.

From a Pod in the same Namespace:

nslookup hostnames
Address 1: kube-dns.kube-system.svc.cluster.local

Name:      hostnames
Address 1: hostnames.default.svc.cluster.local

If this fails, perhaps your Pod and Service are in different Namespaces, try a namespace-qualified name (again, from within a Pod):

nslookup hostnames.default
Address 1: kube-dns.kube-system.svc.cluster.local

Name:      hostnames.default
Address 1: hostnames.default.svc.cluster.local

If this works, you'll need to adjust your app to use a cross-namespace name, or run your app and Service in the same Namespace. If this still fails, try a fully-qualified name:

nslookup hostnames.default.svc.cluster.local
Address 1: kube-dns.kube-system.svc.cluster.local

Name:      hostnames.default.svc.cluster.local
Address 1: hostnames.default.svc.cluster.local

Note the suffix here: "default.svc.cluster.local". The "default" is the Namespace you're operating in. The "svc" denotes that this is a Service. The "cluster.local" is your cluster domain, which COULD be different in your own cluster.

You can also try this from a Node in the cluster:

nslookup hostnames.default.svc.cluster.local

Name:   hostnames.default.svc.cluster.local

If you are able to do a fully-qualified name lookup but not a relative one, you need to check that your /etc/resolv.conf file in your Pod is correct. From within a Pod:

cat /etc/resolv.conf

You should see something like:

search default.svc.cluster.local svc.cluster.local cluster.local
options ndots:5

The nameserver line must indicate your cluster's DNS Service. This is passed into kubelet with the --cluster-dns flag.

The search line must include an appropriate suffix for you to find the Service name. In this case it is looking for Services in the local Namespace ("default.svc.cluster.local"), Services in all Namespaces ("svc.cluster.local"), and lastly for names in the cluster ("cluster.local"). Depending on your own install you might have additional records after that (up to 6 total). The cluster suffix is passed into kubelet with the --cluster-domain flag. Throughout this document, the cluster suffix is assumed to be "cluster.local". Your own clusters might be configured differently, in which case you should change that in all of the previous commands.

The options line must set ndots high enough that your DNS client library considers search paths at all. Kubernetes sets this to 5 by default, which is high enough to cover all of the DNS names it generates.

Does any Service work by DNS name?

If the above still fails, DNS lookups are not working for your Service. You can take a step back and see what else is not working. The Kubernetes master Service should always work. From within a Pod:

nslookup kubernetes.default
Address 1: kube-dns.kube-system.svc.cluster.local

Name:      kubernetes.default
Address 1: kubernetes.default.svc.cluster.local

If this fails, please see the kube-proxy section of this document, or even go back to the top of this document and start over, but instead of debugging your own Service, debug the DNS Service.

Does the Service work by IP?

Assuming you have confirmed that DNS works, the next thing to test is whether your Service works by its IP address. From a Pod in your cluster, access the Service's IP (from kubectl get above).

for i in $(seq 1 3); do 
    wget -qO-

This should produce something like:


If your Service is working, you should get correct responses. If not, there are a number of things that could be going wrong. Read on.

Is the Service defined correctly?

It might sound silly, but you should really double and triple check that your Service is correct and matches your Pod's port. Read back your Service and verify it:

kubectl get service hostnames -o json
    "kind": "Service",
    "apiVersion": "v1",
    "metadata": {
        "name": "hostnames",
        "namespace": "default",
        "uid": "428c8b6c-24bc-11e5-936d-42010af0a9bc",
        "resourceVersion": "347189",
        "creationTimestamp": "2015-07-07T15:24:29Z",
        "labels": {
            "app": "hostnames"
    "spec": {
        "ports": [
                "name": "default",
                "protocol": "TCP",
                "port": 80,
                "targetPort": 9376,
                "nodePort": 0
        "selector": {
            "app": "hostnames"
        "clusterIP": "",
        "type": "ClusterIP",
        "sessionAffinity": "None"
    "status": {
        "loadBalancer": {}
  • Is the Service port you are trying to access listed in spec.ports[]?
  • Is the targetPort correct for your Pods (some Pods use a different port than the Service)?
  • If you meant to use a numeric port, is it a number (9376) or a string "9376"?
  • If you meant to use a named port, do your Pods expose a port with the same name?
  • Is the port's protocol correct for your Pods?

Does the Service have any Endpoints?

If you got this far, you have confirmed that your Service is correctly defined and is resolved by DNS. Now let's check that the Pods you ran are actually being selected by the Service.

Earlier you saw that the Pods were running. You can re-check that:

kubectl get pods -l app=hostnames
NAME                        READY     STATUS    RESTARTS   AGE
hostnames-632524106-bbpiw   1/1       Running   0          1h
hostnames-632524106-ly40y   1/1       Running   0          1h
hostnames-632524106-tlaok   1/1       Running   0          1h

The -l app=hostnames argument is a label selector configured on the Service.

The "AGE" column says that these Pods are about an hour old, which implies that they are running fine and not crashing.

The "RESTARTS" column says that these pods are not crashing frequently or being restarted. Frequent restarts could lead to intermittent connectivity issues. If the restart count is high, read more about how to debug pods.

Inside the Kubernetes system is a control loop which evaluates the selector of every Service and saves the results into a corresponding Endpoints object.

kubectl get endpoints hostnames


This confirms that the endpoints controller has found the correct Pods for your Service. If the ENDPOINTS column is <none>, you should check that the spec.selector field of your Service actually selects for metadata.labels values on your Pods. A common mistake is to have a typo or other error, such as the Service selecting for app=hostnames, but the Deployment specifying run=hostnames, as in versions previous to 1.18, where the kubectl run command could have been also used to create a Deployment.

Are the Pods working?

At this point, you know that your Service exists and has selected your Pods. At the beginning of this walk-through, you verified the Pods themselves. Let's check again that the Pods are actually working - you can bypass the Service mechanism and go straight to the Pods, as listed by the Endpoints above.

From within a Pod:

for ep in; do
    wget -qO- $ep

This should produce something like:


You expect each Pod in the Endpoints list to return its own hostname. If this is not what happens (or whatever the correct behavior is for your own Pods), you should investigate what's happening there.

Is the kube-proxy working?

If you get here, your Service is running, has Endpoints, and your Pods are actually serving. At this point, the whole Service proxy mechanism is suspect. Let's confirm it, piece by piece.

The default implementation of Services, and the one used on most clusters, is kube-proxy. This is a program that runs on every node and configures one of a small set of mechanisms for providing the Service abstraction. If your cluster does not use kube-proxy, the following sections will not apply, and you will have to investigate whatever implementation of Services you are using.

Is kube-proxy running?

Confirm that kube-proxy is running on your Nodes. Running directly on a Node, you should get something like the below:

ps auxw | grep kube-proxy
root  4194  0.4  0.1 101864 17696 ?    Sl Jul04  25:43 /usr/local/bin/kube-proxy --master=https://kubernetes-master --kubeconfig=/var/lib/kube-proxy/kubeconfig --v=2

Next, confirm that it is not failing something obvious, like contacting the master. To do this, you'll have to look at the logs. Accessing the logs depends on your Node OS. On some OSes it is a file, such as /var/log/kube-proxy.log, while other OSes use journalctl to access logs. You should see something like:

I1027 22:14:53.995134    5063 server.go:200] Running in resource-only container "/kube-proxy"
I1027 22:14:53.998163    5063 server.go:247] Using iptables Proxier.
I1027 22:14:53.999055    5063 server.go:255] Tearing down userspace rules. Errors here are acceptable.
I1027 22:14:54.038140    5063 proxier.go:352] Setting endpoints for "kube-system/kube-dns:dns-tcp" to []
I1027 22:14:54.038164    5063 proxier.go:352] Setting endpoints for "kube-system/kube-dns:dns" to []
I1027 22:14:54.038209    5063 proxier.go:352] Setting endpoints for "default/kubernetes:https" to []
I1027 22:14:54.038238    5063 proxier.go:429] Not syncing iptables until Services and Endpoints have been received from master
I1027 22:14:54.040048    5063 proxier.go:294] Adding new service "default/kubernetes:https" at
I1027 22:14:54.040154    5063 proxier.go:294] Adding new service "kube-system/kube-dns:dns" at
I1027 22:14:54.040223    5063 proxier.go:294] Adding new service "kube-system/kube-dns:dns-tcp" at

If you see error messages about not being able to contact the master, you should double-check your Node configuration and installation steps.

One of the possible reasons that kube-proxy cannot run correctly is that the required conntrack binary cannot be found. This may happen on some Linux systems, depending on how you are installing the cluster, for example, you are installing Kubernetes from scratch. If this is the case, you need to manually install the conntrack package (e.g. sudo apt install conntrack on Ubuntu) and then retry.

Kube-proxy can run in one of a few modes. In the log listed above, the line Using iptables Proxier indicates that kube-proxy is running in "iptables" mode. The most common other mode is "ipvs". The older "userspace" mode has largely been replaced by these.

Iptables mode

In "iptables" mode, you should see something like the following on a Node:

iptables-save | grep hostnames
-A KUBE-SEP-57KPRZ3JQVENLNBR -s -m comment --comment "default/hostnames:" -j MARK --set-xmark 0x00004000/0x00004000
-A KUBE-SEP-57KPRZ3JQVENLNBR -p tcp -m comment --comment "default/hostnames:" -m tcp -j DNAT --to-destination
-A KUBE-SEP-WNBA2IHDGP2BOBGZ -s -m comment --comment "default/hostnames:" -j MARK --set-xmark 0x00004000/0x00004000
-A KUBE-SEP-WNBA2IHDGP2BOBGZ -p tcp -m comment --comment "default/hostnames:" -m tcp -j DNAT --to-destination
-A KUBE-SEP-X3P2623AGDH6CDF3 -s -m comment --comment "default/hostnames:" -j MARK --set-xmark 0x00004000/0x00004000
-A KUBE-SEP-X3P2623AGDH6CDF3 -p tcp -m comment --comment "default/hostnames:" -m tcp -j DNAT --to-destination
-A KUBE-SERVICES -d -p tcp -m comment --comment "default/hostnames: cluster IP" -m tcp --dport 80 -j KUBE-SVC-NWV5X2332I4OT4T3
-A KUBE-SVC-NWV5X2332I4OT4T3 -m comment --comment "default/hostnames:" -m statistic --mode random --probability 0.33332999982 -j KUBE-SEP-WNBA2IHDGP2BOBGZ
-A KUBE-SVC-NWV5X2332I4OT4T3 -m comment --comment "default/hostnames:" -m statistic --mode random --probability 0.50000000000 -j KUBE-SEP-X3P2623AGDH6CDF3
-A KUBE-SVC-NWV5X2332I4OT4T3 -m comment --comment "default/hostnames:" -j KUBE-SEP-57KPRZ3JQVENLNBR

For each port of each Service, there should be 1 rule in KUBE-SERVICES and one KUBE-SVC-<hash> chain. For each Pod endpoint, there should be a small number of rules in that KUBE-SVC-<hash> and one KUBE-SEP-<hash> chain with a small number of rules in it. The exact rules will vary based on your exact config (including node-ports and load-balancers).

IPVS mode

In "ipvs" mode, you should see something like the following on a Node:

ipvsadm -ln
Prot LocalAddress:Port Scheduler Flags
  -> RemoteAddress:Port           Forward Weight ActiveConn InActConn
TCP rr
  ->               Masq    1      0          0
  ->               Masq    1      0          0
  ->               Masq    1      0          0

For each port of each Service, plus any NodePorts, external IPs, and load-balancer IPs, kube-proxy will create a virtual server. For each Pod endpoint, it will create corresponding real servers. In this example, service hostnames( has 3 endpoints(,,

Userspace mode

In rare cases, you may be using "userspace" mode. From your Node:

iptables-save | grep hostnames
-A KUBE-PORTALS-CONTAINER -d -p tcp -m comment --comment "default/hostnames:default" -m tcp --dport 80 -j REDIRECT --to-ports 48577
-A KUBE-PORTALS-HOST -d -p tcp -m comment --comment "default/hostnames:default" -m tcp --dport 80 -j DNAT --to-destination

There should be 2 rules for each port of your Service (only one in this example) - a "KUBE-PORTALS-CONTAINER" and a "KUBE-PORTALS-HOST".

Almost nobody should be using the "userspace" mode any more, so you won't spend more time on it here.

Is kube-proxy proxying?

Assuming you do see one the above cases, try again to access your Service by IP from one of your Nodes:


If this fails and you are using the userspace proxy, you can try accessing the proxy directly. If you are using the iptables proxy, skip this section.

Look back at the iptables-save output above, and extract the port number that kube-proxy is using for your Service. In the above examples it is "48577". Now connect to that:

curl localhost:48577

If this still fails, look at the kube-proxy logs for specific lines like:

Setting endpoints for default/hostnames:default to []

If you don't see those, try restarting kube-proxy with the -v flag set to 4, and then look at the logs again.

Edge case: A Pod fails to reach itself via the Service IP

This might sound unlikely, but it does happen and it is supposed to work.

This can happen when the network is not properly configured for "hairpin" traffic, usually when kube-proxy is running in iptables mode and Pods are connected with bridge network. The Kubelet exposes a hairpin-mode flag that allows endpoints of a Service to loadbalance back to themselves if they try to access their own Service VIP. The hairpin-mode flag must either be set to hairpin-veth or promiscuous-bridge.

The common steps to trouble shoot this are as follows:

  • Confirm hairpin-mode is set to hairpin-veth or promiscuous-bridge. You should see something like the below. hairpin-mode is set to promiscuous-bridge in the following example.
ps auxw | grep kubelet
root      3392  1.1  0.8 186804 65208 ?        Sl   00:51  11:11 /usr/local/bin/kubelet --enable-debugging-handlers=true --config=/etc/kubernetes/manifests --allow-privileged=True --v=4 --cluster-dns= --cluster-domain=cluster.local --configure-cbr0=true --cgroup-root=/ --system-cgroups=/system --hairpin-mode=promiscuous-bridge --runtime-cgroups=/docker-daemon --kubelet-cgroups=/kubelet --babysit-daemons=true --max-pods=110 --serialize-image-pulls=false --outofdisk-transition-frequency=0
  • Confirm the effective hairpin-mode. To do this, you'll have to look at kubelet log. Accessing the logs depends on your Node OS. On some OSes it is a file, such as /var/log/kubelet.log, while other OSes use journalctl to access logs. Please be noted that the effective hairpin mode may not match --hairpin-mode flag due to compatibility. Check if there is any log lines with key word hairpin in kubelet.log. There should be log lines indicating the effective hairpin mode, like something below.
I0629 00:51:43.648698    3252 kubelet.go:380] Hairpin mode set to "promiscuous-bridge"
  • If the effective hairpin mode is hairpin-veth, ensure the Kubelet has the permission to operate in /sys on node. If everything works properly, you should see something like:
for intf in /sys/devices/virtual/net/cbr0/brif/*; do cat $intf/hairpin_mode; done
  • If the effective hairpin mode is promiscuous-bridge, ensure Kubelet has the permission to manipulate linux bridge on node. If cbr0 bridge is used and configured properly, you should see:
ifconfig cbr0 |grep PROMISC
  • Seek help if none of above works out.

Seek help

If you get this far, something very strange is happening. Your Service is running, has Endpoints, and your Pods are actually serving. You have DNS working, and kube-proxy does not seem to be misbehaving. And yet your Service is not working. Please let us know what is going on, so we can help investigate!

Contact us on Slack or Forum or GitHub.

What's next

Visit the troubleshooting overview document for more information.

1.3 - Debug a StatefulSet

This task shows you how to debug a StatefulSet.

Before you begin

  • You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster.
  • You should have a StatefulSet running that you want to investigate.

Debugging a StatefulSet

In order to list all the pods which belong to a StatefulSet, which have a label app=myapp set on them, you can use the following:

kubectl get pods -l app=myapp

If you find that any Pods listed are in Unknown or Terminating state for an extended period of time, refer to the Deleting StatefulSet Pods task for instructions on how to deal with them. You can debug individual Pods in a StatefulSet using the Debugging Pods guide.

What's next

Learn more about debugging an init-container.

1.4 - Debug Init Containers

This page shows how to investigate problems related to the execution of Init Containers. The example command lines below refer to the Pod as <pod-name> and the Init Containers as <init-container-1> and <init-container-2>.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Checking the status of Init Containers

Display the status of your pod:

kubectl get pod <pod-name>

For example, a status of Init:1/2 indicates that one of two Init Containers has completed successfully:

<pod-name>   0/1       Init:1/2   0          7s

See Understanding Pod status for more examples of status values and their meanings.

Getting details about Init Containers

View more detailed information about Init Container execution:

kubectl describe pod <pod-name>

For example, a Pod with two Init Containers might show the following:

Init Containers:
    Container ID:    ...
    State:           Terminated
      Reason:        Completed
      Exit Code:     0
      Started:       ...
      Finished:      ...
    Ready:           True
    Restart Count:   0
    Container ID:    ...
    State:           Waiting
      Reason:        CrashLoopBackOff
    Last State:      Terminated
      Reason:        Error
      Exit Code:     1
      Started:       ...
      Finished:      ...
    Ready:           False
    Restart Count:   3

You can also access the Init Container statuses programmatically by reading the status.initContainerStatuses field on the Pod Spec:

kubectl get pod nginx --template '{{.status.initContainerStatuses}}'

This command will return the same information as above in raw JSON.

Accessing logs from Init Containers

Pass the Init Container name along with the Pod name to access its logs.

kubectl logs <pod-name> -c <init-container-2>

Init Containers that run a shell script print commands as they're executed. For example, you can do this in Bash by running set -x at the beginning of the script.

Understanding Pod status

A Pod status beginning with Init: summarizes the status of Init Container execution. The table below describes some example status values that you might see while debugging Init Containers.

Init:N/MThe Pod has M Init Containers, and N have completed so far.
Init:ErrorAn Init Container has failed to execute.
Init:CrashLoopBackOffAn Init Container has failed repeatedly.
PendingThe Pod has not yet begun executing Init Containers.
PodInitializing or RunningThe Pod has already finished executing Init Containers.

1.5 - Debug Running Pods

This page explains how to debug Pods running (or crashing) on a Node.

Before you begin

  • Your Pod should already be scheduled and running. If your Pod is not yet running, start with Debugging Pods.
  • For some of the advanced debugging steps you need to know on which Node the Pod is running and have shell access to run commands on that Node. You don't need that access to run the standard debug steps that use kubectl.

Using kubectl describe pod to fetch details about pods

For this example we'll use a Deployment to create two pods, similar to the earlier example.

apiVersion: apps/v1
kind: Deployment
  name: nginx-deployment
      app: nginx
  replicas: 2
        app: nginx
      - name: nginx
        image: nginx
            memory: "128Mi"
            cpu: "500m"
        - containerPort: 80

Create deployment by running following command:

kubectl apply -f
deployment.apps/nginx-deployment created

Check pod status by following command:

kubectl get pods
NAME                                READY   STATUS    RESTARTS   AGE
nginx-deployment-67d4bdd6f5-cx2nz   1/1     Running   0          13s
nginx-deployment-67d4bdd6f5-w6kd7   1/1     Running   0          13s

We can retrieve a lot more information about each of these pods using kubectl describe pod. For example:

kubectl describe pod nginx-deployment-67d4bdd6f5-w6kd7
Name:         nginx-deployment-67d4bdd6f5-w6kd7
Namespace:    default
Priority:     0
Node:         kube-worker-1/
Start Time:   Thu, 17 Feb 2022 16:51:01 -0500
Labels:       app=nginx
Annotations:  <none>
Status:       Running
  IP:           2001:db8::1
Controlled By:  ReplicaSet/nginx-deployment-67d4bdd6f5
    Container ID:   containerd://5403af59a2b46ee5a23fb0ae4b1e077f7ca5c5fb7af16e1ab21c00e0e616462a
    Image:          nginx
    Image ID:
    Port:           80/TCP
    Host Port:      0/TCP
    State:          Running
      Started:      Thu, 17 Feb 2022 16:51:05 -0500
    Ready:          True
    Restart Count:  0
      cpu:     500m
      memory:  128Mi
      cpu:        500m
      memory:     128Mi
    Environment:  <none>
      /var/run/secrets/ from kube-api-access-bgsgp (ro)
  Type              Status
  Initialized       True 
  Ready             True 
  ContainersReady   True 
  PodScheduled      True 
    Type:                    Projected (a volume that contains injected data from multiple sources)
    TokenExpirationSeconds:  3607
    ConfigMapName:           kube-root-ca.crt
    ConfigMapOptional:       <nil>
    DownwardAPI:             true
QoS Class:                   Guaranteed
Node-Selectors:              <none>
Tolerations:        op=Exists for 300s
                    op=Exists for 300s
  Type    Reason     Age   From               Message
  ----    ------     ----  ----               -------
  Normal  Scheduled  34s   default-scheduler  Successfully assigned default/nginx-deployment-67d4bdd6f5-w6kd7 to kube-worker-1
  Normal  Pulling    31s   kubelet            Pulling image "nginx"
  Normal  Pulled     30s   kubelet            Successfully pulled image "nginx" in 1.146417389s
  Normal  Created    30s   kubelet            Created container nginx
  Normal  Started    30s   kubelet            Started container nginx

Here you can see configuration information about the container(s) and Pod (labels, resource requirements, etc.), as well as status information about the container(s) and Pod (state, readiness, restart count, events, etc.).

The container state is one of Waiting, Running, or Terminated. Depending on the state, additional information will be provided -- here you can see that for a container in Running state, the system tells you when the container started.

Ready tells you whether the container passed its last readiness probe. (In this case, the container does not have a readiness probe configured; the container is assumed to be ready if no readiness probe is configured.)

Restart Count tells you how many times the container has been restarted; this information can be useful for detecting crash loops in containers that are configured with a restart policy of 'always.'

Currently the only Condition associated with a Pod is the binary Ready condition, which indicates that the pod is able to service requests and should be added to the load balancing pools of all matching services.

Lastly, you see a log of recent events related to your Pod. The system compresses multiple identical events by indicating the first and last time it was seen and the number of times it was seen. "From" indicates the component that is logging the event, "SubobjectPath" tells you which object (e.g. container within the pod) is being referred to, and "Reason" and "Message" tell you what happened.

Example: debugging Pending Pods

A common scenario that you can detect using events is when you've created a Pod that won't fit on any node. For example, the Pod might request more resources than are free on any node, or it might specify a label selector that doesn't match any nodes. Let's say we created the previous Deployment with 5 replicas (instead of 2) and requesting 600 millicores instead of 500, on a four-node cluster where each (virtual) machine has 1 CPU. In that case one of the Pods will not be able to schedule. (Note that because of the cluster addon pods such as fluentd, skydns, etc., that run on each node, if we requested 1000 millicores then none of the Pods would be able to schedule.)

kubectl get pods
NAME                                READY     STATUS    RESTARTS   AGE
nginx-deployment-1006230814-6winp   1/1       Running   0          7m
nginx-deployment-1006230814-fmgu3   1/1       Running   0          7m
nginx-deployment-1370807587-6ekbw   1/1       Running   0          1m
nginx-deployment-1370807587-fg172   0/1       Pending   0          1m
nginx-deployment-1370807587-fz9sd   0/1       Pending   0          1m

To find out why the nginx-deployment-1370807587-fz9sd pod is not running, we can use kubectl describe pod on the pending Pod and look at its events:

kubectl describe pod nginx-deployment-1370807587-fz9sd
  Name:		nginx-deployment-1370807587-fz9sd
  Namespace:	default
  Node:		/
  Labels:		app=nginx,pod-template-hash=1370807587
  Status:		Pending
  Controllers:	ReplicaSet/nginx-deployment-1370807587
      Image:	nginx
      Port:	80/TCP
      QoS Tier:
        memory:	Guaranteed
        cpu:	Guaranteed
        cpu:	1
        memory:	128Mi
        cpu:	1
        memory:	128Mi
      Environment Variables:
      Type:	Secret (a volume populated by a Secret)
      SecretName:	default-token-4bcbi
    FirstSeen	LastSeen	Count	From			        SubobjectPath	Type		Reason			    Message
    ---------	--------	-----	----			        -------------	--------	------			    -------
    1m		    48s		    7	    {default-scheduler }			        Warning		FailedScheduling	pod (nginx-deployment-1370807587-fz9sd) failed to fit in any node
  fit failure on node (kubernetes-node-6ta5): Node didn't have enough resource: CPU, requested: 1000, used: 1420, capacity: 2000
  fit failure on node (kubernetes-node-wul5): Node didn't have enough resource: CPU, requested: 1000, used: 1100, capacity: 2000

Here you can see the event generated by the scheduler saying that the Pod failed to schedule for reason FailedScheduling (and possibly others). The message tells us that there were not enough resources for the Pod on any of the nodes.

To correct this situation, you can use kubectl scale to update your Deployment to specify four or fewer replicas. (Or you could leave the one Pod pending, which is harmless.)

Events such as the ones you saw at the end of kubectl describe pod are persisted in etcd and provide high-level information on what is happening in the cluster. To list all events you can use

kubectl get events

but you have to remember that events are namespaced. This means that if you're interested in events for some namespaced object (e.g. what happened with Pods in namespace my-namespace) you need to explicitly provide a namespace to the command:

kubectl get events --namespace=my-namespace

To see events from all namespaces, you can use the --all-namespaces argument.

In addition to kubectl describe pod, another way to get extra information about a pod (beyond what is provided by kubectl get pod) is to pass the -o yaml output format flag to kubectl get pod. This will give you, in YAML format, even more information than kubectl describe pod--essentially all of the information the system has about the Pod. Here you will see things like annotations (which are key-value metadata without the label restrictions, that is used internally by Kubernetes system components), restart policy, ports, and volumes.

kubectl get pod nginx-deployment-1006230814-6winp -o yaml
apiVersion: v1
kind: Pod
  creationTimestamp: "2022-02-17T21:51:01Z"
  generateName: nginx-deployment-67d4bdd6f5-
    app: nginx
    pod-template-hash: 67d4bdd6f5
  name: nginx-deployment-67d4bdd6f5-w6kd7
  namespace: default
  - apiVersion: apps/v1
    blockOwnerDeletion: true
    controller: true
    kind: ReplicaSet
    name: nginx-deployment-67d4bdd6f5
    uid: 7d41dfd4-84c0-4be4-88ab-cedbe626ad82
  resourceVersion: "1364"
  uid: a6501da1-0447-4262-98eb-c03d4002222e
  - image: nginx
    imagePullPolicy: Always
    name: nginx
    - containerPort: 80
      protocol: TCP
        cpu: 500m
        memory: 128Mi
        cpu: 500m
        memory: 128Mi
    terminationMessagePath: /dev/termination-log
    terminationMessagePolicy: File
    - mountPath: /var/run/secrets/
      name: kube-api-access-bgsgp
      readOnly: true
  dnsPolicy: ClusterFirst
  enableServiceLinks: true
  nodeName: kube-worker-1
  preemptionPolicy: PreemptLowerPriority
  priority: 0
  restartPolicy: Always
  schedulerName: default-scheduler
  securityContext: {}
  serviceAccount: default
  serviceAccountName: default
  terminationGracePeriodSeconds: 30
  - effect: NoExecute
    operator: Exists
    tolerationSeconds: 300
  - effect: NoExecute
    operator: Exists
    tolerationSeconds: 300
  - name: kube-api-access-bgsgp
      defaultMode: 420
      - serviceAccountToken:
          expirationSeconds: 3607
          path: token
      - configMap:
          - key: ca.crt
            path: ca.crt
          name: kube-root-ca.crt
      - downwardAPI:
          - fieldRef:
              apiVersion: v1
              fieldPath: metadata.namespace
            path: namespace
  - lastProbeTime: null
    lastTransitionTime: "2022-02-17T21:51:01Z"
    status: "True"
    type: Initialized
  - lastProbeTime: null
    lastTransitionTime: "2022-02-17T21:51:06Z"
    status: "True"
    type: Ready
  - lastProbeTime: null
    lastTransitionTime: "2022-02-17T21:51:06Z"
    status: "True"
    type: ContainersReady
  - lastProbeTime: null
    lastTransitionTime: "2022-02-17T21:51:01Z"
    status: "True"
    type: PodScheduled
  - containerID: containerd://5403af59a2b46ee5a23fb0ae4b1e077f7ca5c5fb7af16e1ab21c00e0e616462a
    lastState: {}
    name: nginx
    ready: true
    restartCount: 0
    started: true
        startedAt: "2022-02-17T21:51:05Z"
  phase: Running
  - ip:
  - ip: 2001:db8::1
  qosClass: Guaranteed
  startTime: "2022-02-17T21:51:01Z"

Examining pod logs

First, look at the logs of the affected container:

kubectl logs ${POD_NAME} ${CONTAINER_NAME}

If your container has previously crashed, you can access the previous container's crash log with:

kubectl logs --previous ${POD_NAME} ${CONTAINER_NAME}

Debugging with container exec

If the container image includes debugging utilities, as is the case with images built from Linux and Windows OS base images, you can run commands inside a specific container with kubectl exec:

kubectl exec ${POD_NAME} -c ${CONTAINER_NAME} -- ${CMD} ${ARG1} ${ARG2} ... ${ARGN}

As an example, to look at the logs from a running Cassandra pod, you might run

kubectl exec cassandra -- cat /var/log/cassandra/system.log

You can run a shell that's connected to your terminal using the -i and -t arguments to kubectl exec, for example:

kubectl exec -it cassandra -- sh

For more details, see Get a Shell to a Running Container.

Debugging with an ephemeral debug container

FEATURE STATE: Kubernetes v1.23 [beta]

Ephemeral containers are useful for interactive troubleshooting when kubectl exec is insufficient because a container has crashed or a container image doesn't include debugging utilities, such as with distroless images.

Example debugging using ephemeral containers

You can use the kubectl debug command to add ephemeral containers to a running Pod. First, create a pod for the example:

kubectl run ephemeral-demo --restart=Never

The examples in this section use the pause container image because it does not contain debugging utilities, but this method works with all container images.

If you attempt to use kubectl exec to create a shell you will see an error because there is no shell in this container image.

kubectl exec -it ephemeral-demo -- sh
OCI runtime exec failed: exec failed: container_linux.go:346: starting container process caused "exec: \"sh\": executable file not found in $PATH": unknown

You can instead add a debugging container using kubectl debug. If you specify the -i/--interactive argument, kubectl will automatically attach to the console of the Ephemeral Container.

kubectl debug -it ephemeral-demo --image=busybox:1.28 --target=ephemeral-demo
Defaulting debug container name to debugger-8xzrl.
If you don't see a command prompt, try pressing enter.
/ #

This command adds a new busybox container and attaches to it. The --target parameter targets the process namespace of another container. It's necessary here because kubectl run does not enable process namespace sharing in the pod it creates.

You can view the state of the newly created ephemeral container using kubectl describe:

kubectl describe pod ephemeral-demo
Ephemeral Containers:
    Container ID:   docker://b888f9adfd15bd5739fefaa39e1df4dd3c617b9902082b1cfdc29c4028ffb2eb
    Image:          busybox
    Image ID:       docker-pullable://busybox@sha256:1828edd60c5efd34b2bf5dd3282ec0cc04d47b2ff9caa0b6d4f07a21d1c08084
    Port:           <none>
    Host Port:      <none>
    State:          Running
      Started:      Wed, 12 Feb 2020 14:25:42 +0100
    Ready:          False
    Restart Count:  0
    Environment:    <none>
    Mounts:         <none>

Use kubectl delete to remove the Pod when you're finished:

kubectl delete pod ephemeral-demo

Debugging using a copy of the Pod

Sometimes Pod configuration options make it difficult to troubleshoot in certain situations. For example, you can't run kubectl exec to troubleshoot your container if your container image does not include a shell or if your application crashes on startup. In these situations you can use kubectl debug to create a copy of the Pod with configuration values changed to aid debugging.

Copying a Pod while adding a new container

Adding a new container can be useful when your application is running but not behaving as you expect and you'd like to add additional troubleshooting utilities to the Pod.

For example, maybe your application's container images are built on busybox but you need debugging utilities not included in busybox. You can simulate this scenario using kubectl run:

kubectl run myapp --image=busybox:1.28 --restart=Never -- sleep 1d

Run this command to create a copy of myapp named myapp-debug that adds a new Ubuntu container for debugging:

kubectl debug myapp -it --image=ubuntu --share-processes --copy-to=myapp-debug
Defaulting debug container name to debugger-w7xmf.
If you don't see a command prompt, try pressing enter.

Don't forget to clean up the debugging Pod when you're finished with it:

kubectl delete pod myapp myapp-debug

Copying a Pod while changing its command

Sometimes it's useful to change the command for a container, for example to add a debugging flag or because the application is crashing.

To simulate a crashing application, use kubectl run to create a container that immediately exits:

kubectl run --image=busybox:1.28 myapp -- false

You can see using kubectl describe pod myapp that this container is crashing:

    Image:         busybox
    State:          Waiting
      Reason:       CrashLoopBackOff
    Last State:     Terminated
      Reason:       Error
      Exit Code:    1

You can use kubectl debug to create a copy of this Pod with the command changed to an interactive shell:

kubectl debug myapp -it --copy-to=myapp-debug --container=myapp -- sh
If you don't see a command prompt, try pressing enter.
/ #

Now you have an interactive shell that you can use to perform tasks like checking filesystem paths or running the container command manually.

Don't forget to clean up the debugging Pod when you're finished with it:

kubectl delete pod myapp myapp-debug

Copying a Pod while changing container images

In some situations you may want to change a misbehaving Pod from its normal production container images to an image containing a debugging build or additional utilities.

As an example, create a Pod using kubectl run:

kubectl run myapp --image=busybox:1.28 --restart=Never -- sleep 1d

Now use kubectl debug to make a copy and change its container image to ubuntu:

kubectl debug myapp --copy-to=myapp-debug --set-image=*=ubuntu

The syntax of --set-image uses the same container_name=image syntax as kubectl set image. *=ubuntu means change the image of all containers to ubuntu.

Don't forget to clean up the debugging Pod when you're finished with it:

kubectl delete pod myapp myapp-debug

Debugging via a shell on the node

If none of these approaches work, you can find the Node on which the Pod is running and create a privileged Pod running in the host namespaces. To create an interactive shell on a node using kubectl debug, run:

kubectl debug node/mynode -it --image=ubuntu
Creating debugging pod node-debugger-mynode-pdx84 with container debugger on node mynode.
If you don't see a command prompt, try pressing enter.

When creating a debugging session on a node, keep in mind that:

  • kubectl debug automatically generates the name of the new Pod based on the name of the Node.
  • The container runs in the host IPC, Network, and PID namespaces.
  • The root filesystem of the Node will be mounted at /host.

Don't forget to clean up the debugging Pod when you're finished with it:

kubectl delete pod node-debugger-mynode-pdx84

1.6 - Determine the Reason for Pod Failure

This page shows how to write and read a Container termination message.

Termination messages provide a way for containers to write information about fatal events to a location where it can be easily retrieved and surfaced by tools like dashboards and monitoring software. In most cases, information that you put in a termination message should also be written to the general Kubernetes logs.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Writing and reading a termination message

In this exercise, you create a Pod that runs one container. The configuration file specifies a command that runs when the container starts.

apiVersion: v1
kind: Pod
  name: termination-demo
  - name: termination-demo-container
    image: debian
    command: ["/bin/sh"]
    args: ["-c", "sleep 10 && echo Sleep expired > /dev/termination-log"]
  1. Create a Pod based on the YAML configuration file:

     kubectl apply -f

    In the YAML file, in the command and args fields, you can see that the container sleeps for 10 seconds and then writes "Sleep expired" to the /dev/termination-log file. After the container writes the "Sleep expired" message, it terminates.

  2. Display information about the Pod:

     kubectl get pod termination-demo

    Repeat the preceding command until the Pod is no longer running.

  3. Display detailed information about the Pod:

     kubectl get pod termination-demo --output=yaml

    The output includes the "Sleep expired" message:

     apiVersion: v1
     kind: Pod
             containerID: ...
             exitCode: 0
             finishedAt: ...
             message: |
               Sleep expired
  4. Use a Go template to filter the output so that it includes only the termination message:

     kubectl get pod termination-demo -o go-template="{{range .status.containerStatuses}}{{.lastState.terminated.message}}{{end}}"

If you are running a multi-container pod, you can use a Go template to include the container's name. By doing so, you can discover which of the containers is failing:

kubectl get pod multi-container-pod -o go-template='{{range .status.containerStatuses}}{{printf "%s:\n%s\n\n" .name .lastState.terminated.message}}{{end}}'

Customizing the termination message

Kubernetes retrieves termination messages from the termination message file specified in the terminationMessagePath field of a Container, which has a default value of /dev/termination-log. By customizing this field, you can tell Kubernetes to use a different file. Kubernetes use the contents from the specified file to populate the Container's status message on both success and failure.

The termination message is intended to be brief final status, such as an assertion failure message. The kubelet truncates messages that are longer than 4096 bytes. The total message length across all containers will be limited to 12KiB. The default termination message path is /dev/termination-log. You cannot set the termination message path after a Pod is launched

In the following example, the container writes termination messages to /tmp/my-log for Kubernetes to retrieve:

apiVersion: v1
kind: Pod
  name: msg-path-demo
  - name: msg-path-demo-container
    image: debian
    terminationMessagePath: "/tmp/my-log"

Moreover, users can set the terminationMessagePolicy field of a Container for further customization. This field defaults to "File" which means the termination messages are retrieved only from the termination message file. By setting the terminationMessagePolicy to "FallbackToLogsOnError", you can tell Kubernetes to use the last chunk of container log output if the termination message file is empty and the container exited with an error. The log output is limited to 2048 bytes or 80 lines, whichever is smaller.

What's next

1.7 - Get a Shell to a Running Container

This page shows how to use kubectl exec to get a shell to a running container.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

Getting a shell to a container

In this exercise, you create a Pod that has one container. The container runs the nginx image. Here is the configuration file for the Pod:

apiVersion: v1
kind: Pod
  name: shell-demo
  - name: shared-data
    emptyDir: {}
  - name: nginx
    image: nginx
    - name: shared-data
      mountPath: /usr/share/nginx/html
  hostNetwork: true
  dnsPolicy: Default

Create the Pod:

kubectl apply -f

Verify that the container is running:

kubectl get pod shell-demo

Get a shell to the running container:

kubectl exec --stdin --tty shell-demo -- /bin/bash

In your shell, list the root directory:

# Run this inside the container
ls /

In your shell, experiment with other commands. Here are some examples:

# You can run these example commands inside the container
ls /
cat /proc/mounts
cat /proc/1/maps
apt-get update
apt-get install -y tcpdump
apt-get install -y lsof
apt-get install -y procps
ps aux
ps aux | grep nginx

Writing the root page for nginx

Look again at the configuration file for your Pod. The Pod has an emptyDir volume, and the container mounts the volume at /usr/share/nginx/html.

In your shell, create an index.html file in the /usr/share/nginx/html directory:

# Run this inside the container
echo 'Hello shell demo' > /usr/share/nginx/html/index.html

In your shell, send a GET request to the nginx server:

# Run this in the shell inside your container
apt-get update
apt-get install curl
curl http://localhost/

The output shows the text that you wrote to the index.html file:

Hello shell demo

When you are finished with your shell, enter exit.

exit # To quit the shell in the container

Running individual commands in a container

In an ordinary command window, not your shell, list the environment variables in the running container:

kubectl exec shell-demo env

Experiment with running other commands. Here are some examples:

kubectl exec shell-demo -- ps aux
kubectl exec shell-demo -- ls /
kubectl exec shell-demo -- cat /proc/1/mounts

Opening a shell when a Pod has more than one container

If a Pod has more than one container, use --container or -c to specify a container in the kubectl exec command. For example, suppose you have a Pod named my-pod, and the Pod has two containers named main-app and helper-app. The following command would open a shell to the main-app container.

kubectl exec -i -t my-pod --container main-app -- /bin/bash

What's next

2 - Troubleshooting Clusters

Debugging common cluster issues.

This doc is about cluster troubleshooting; we assume you have already ruled out your application as the root cause of the problem you are experiencing. See the application troubleshooting guide for tips on application debugging. You may also visit the troubleshooting overview document for more information.

Listing your cluster

The first thing to debug in your cluster is if your nodes are all registered correctly.

Run the following command:

kubectl get nodes

And verify that all of the nodes you expect to see are present and that they are all in the Ready state.

To get detailed information about the overall health of your cluster, you can run:

kubectl cluster-info dump

Example: debugging a down/unreachable node

Sometimes when debugging it can be useful to look at the status of a node -- for example, because you've noticed strange behavior of a Pod that's running on the node, or to find out why a Pod won't schedule onto the node. As with Pods, you can use kubectl describe node and kubectl get node -o yaml to retrieve detailed information about nodes. For example, here's what you'll see if a node is down (disconnected from the network, or kubelet dies and won't restart, etc.). Notice the events that show the node is NotReady, and also notice that the pods are no longer running (they are evicted after five minutes of NotReady status).

kubectl get nodes
NAME                     STATUS       ROLES     AGE     VERSION
kube-worker-1            NotReady     <none>    1h      v1.23.3
kubernetes-node-bols     Ready        <none>    1h      v1.23.3
kubernetes-node-st6x     Ready        <none>    1h      v1.23.3
kubernetes-node-unaj     Ready        <none>    1h      v1.23.3
kubectl describe node kube-worker-1
Name:               kube-worker-1
Roles:              <none>
Annotations: /run/containerd/containerd.sock
CreationTimestamp:  Thu, 17 Feb 2022 16:46:30 -0500
Unschedulable:      false
  HolderIdentity:  kube-worker-1
  AcquireTime:     <unset>
  RenewTime:       Thu, 17 Feb 2022 17:13:09 -0500
  Type                 Status    LastHeartbeatTime                 LastTransitionTime                Reason              Message
  ----                 ------    -----------------                 ------------------                ------              -------
  NetworkUnavailable   False     Thu, 17 Feb 2022 17:09:13 -0500   Thu, 17 Feb 2022 17:09:13 -0500   WeaveIsUp           Weave pod has set this
  MemoryPressure       Unknown   Thu, 17 Feb 2022 17:12:40 -0500   Thu, 17 Feb 2022 17:13:52 -0500   NodeStatusUnknown   Kubelet stopped posting node status.
  DiskPressure         Unknown   Thu, 17 Feb 2022 17:12:40 -0500   Thu, 17 Feb 2022 17:13:52 -0500   NodeStatusUnknown   Kubelet stopped posting node status.
  PIDPressure          Unknown   Thu, 17 Feb 2022 17:12:40 -0500   Thu, 17 Feb 2022 17:13:52 -0500   NodeStatusUnknown   Kubelet stopped posting node status.
  Ready                Unknown   Thu, 17 Feb 2022 17:12:40 -0500   Thu, 17 Feb 2022 17:13:52 -0500   NodeStatusUnknown   Kubelet stopped posting node status.
  Hostname:    kube-worker-1
  cpu:                2
  ephemeral-storage:  15372232Ki
  hugepages-2Mi:      0
  memory:             2025188Ki
  pods:               110
  cpu:                2
  ephemeral-storage:  14167048988
  hugepages-2Mi:      0
  memory:             1922788Ki
  pods:               110
System Info:
  Machine ID:                 9384e2927f544209b5d7b67474bbf92b
  System UUID:                aa829ca9-73d7-064d-9019-df07404ad448
  Boot ID:                    5a295a03-aaca-4340-af20-1327fa5dab5c
  Kernel Version:             5.13.0-28-generic
  OS Image:                   Ubuntu 21.10
  Operating System:           linux
  Architecture:               amd64
  Container Runtime Version:  containerd://1.5.9
  Kubelet Version:            v1.23.3
  Kube-Proxy Version:         v1.23.3
Non-terminated Pods:          (4 in total)
  Namespace                   Name                                 CPU Requests  CPU Limits  Memory Requests  Memory Limits  Age
  ---------                   ----                                 ------------  ----------  ---------------  -------------  ---
  default                     nginx-deployment-67d4bdd6f5-cx2nz    500m (25%)    500m (25%)  128Mi (6%)       128Mi (6%)     23m
  default                     nginx-deployment-67d4bdd6f5-w6kd7    500m (25%)    500m (25%)  128Mi (6%)       128Mi (6%)     23m
  kube-system                 kube-proxy-dnxbz                     0 (0%)        0 (0%)      0 (0%)           0 (0%)         28m
  kube-system                 weave-net-gjxxp                      100m (5%)     0 (0%)      200Mi (10%)      0 (0%)         28m
Allocated resources:
  (Total limits may be over 100 percent, i.e., overcommitted.)
  Resource           Requests     Limits
  --------           --------     ------
  cpu                1100m (55%)  1 (50%)
  memory             456Mi (24%)  256Mi (13%)
  ephemeral-storage  0 (0%)       0 (0%)
  hugepages-2Mi      0 (0%)       0 (0%)
kubectl get node kube-worker-1 -o yaml
apiVersion: v1
kind: Node
  annotations: /run/containerd/containerd.sock "0" "true"
  creationTimestamp: "2022-02-17T21:46:30Z"
  labels: amd64 linux amd64 kube-worker-1 linux
  name: kube-worker-1
  resourceVersion: "4026"
  uid: 98efe7cb-2978-4a0b-842a-1a7bf12c05f8
spec: {}
  - address:
    type: InternalIP
  - address: kube-worker-1
    type: Hostname
    cpu: "2"
    ephemeral-storage: "14167048988"
    hugepages-2Mi: "0"
    memory: 1922788Ki
    pods: "110"
    cpu: "2"
    ephemeral-storage: 15372232Ki
    hugepages-2Mi: "0"
    memory: 2025188Ki
    pods: "110"
  - lastHeartbeatTime: "2022-02-17T22:20:32Z"
    lastTransitionTime: "2022-02-17T22:20:32Z"
    message: Weave pod has set this
    reason: WeaveIsUp
    status: "False"
    type: NetworkUnavailable
  - lastHeartbeatTime: "2022-02-17T22:20:15Z"
    lastTransitionTime: "2022-02-17T22:13:25Z"
    message: kubelet has sufficient memory available
    reason: KubeletHasSufficientMemory
    status: "False"
    type: MemoryPressure
  - lastHeartbeatTime: "2022-02-17T22:20:15Z"
    lastTransitionTime: "2022-02-17T22:13:25Z"
    message: kubelet has no disk pressure
    reason: KubeletHasNoDiskPressure
    status: "False"
    type: DiskPressure
  - lastHeartbeatTime: "2022-02-17T22:20:15Z"
    lastTransitionTime: "2022-02-17T22:13:25Z"
    message: kubelet has sufficient PID available
    reason: KubeletHasSufficientPID
    status: "False"
    type: PIDPressure
  - lastHeartbeatTime: "2022-02-17T22:20:15Z"
    lastTransitionTime: "2022-02-17T22:15:15Z"
    message: kubelet is posting ready status. AppArmor enabled
    reason: KubeletReady
    status: "True"
    type: Ready
      Port: 10250
    architecture: amd64
    bootID: 22333234-7a6b-44d4-9ce1-67e31dc7e369
    containerRuntimeVersion: containerd://1.5.9
    kernelVersion: 5.13.0-28-generic
    kubeProxyVersion: v1.23.3
    kubeletVersion: v1.23.3
    machineID: 9384e2927f544209b5d7b67474bbf92b
    operatingSystem: linux
    osImage: Ubuntu 21.10
    systemUUID: aa829ca9-73d7-064d-9019-df07404ad448

Looking at logs

For now, digging deeper into the cluster requires logging into the relevant machines. Here are the locations of the relevant log files. On systemd-based systems, you may need to use journalctl instead of examining log files.

Control Plane nodes

  • /var/log/kube-apiserver.log - API Server, responsible for serving the API
  • /var/log/kube-scheduler.log - Scheduler, responsible for making scheduling decisions
  • /var/log/kube-controller-manager.log - a component that runs most Kubernetes built-in controllers, with the notable exception of scheduling (the kube-scheduler handles scheduling).

Worker Nodes

  • /var/log/kubelet.log - logs from the kubelet, responsible for running containers on the node
  • /var/log/kube-proxy.log - logs from kube-proxy, which is responsible for directing traffic to Service endpoints

Cluster failure modes

This is an incomplete list of things that could go wrong, and how to adjust your cluster setup to mitigate the problems.

Contributing causes

  • VM(s) shutdown
  • Network partition within cluster, or between cluster and users
  • Crashes in Kubernetes software
  • Data loss or unavailability of persistent storage (e.g. GCE PD or AWS EBS volume)
  • Operator error, for example misconfigured Kubernetes software or application software

Specific scenarios

  • API server VM shutdown or apiserver crashing
    • Results
      • unable to stop, update, or start new pods, services, replication controller
      • existing pods and services should continue to work normally, unless they depend on the Kubernetes API
  • API server backing storage lost
    • Results
      • the kube-apiserver component fails to start successfully and become healthy
      • kubelets will not be able to reach it but will continue to run the same pods and provide the same service proxying
      • manual recovery or recreation of apiserver state necessary before apiserver is restarted
  • Supporting services (node controller, replication controller manager, scheduler, etc) VM shutdown or crashes
    • currently those are colocated with the apiserver, and their unavailability has similar consequences as apiserver
    • in future, these will be replicated as well and may not be co-located
    • they do not have their own persistent state
  • Individual node (VM or physical machine) shuts down
    • Results
      • pods on that Node stop running
  • Network partition
    • Results
      • partition A thinks the nodes in partition B are down; partition B thinks the apiserver is down. (Assuming the master VM ends up in partition A.)
  • Kubelet software fault
    • Results
      • crashing kubelet cannot start new pods on the node
      • kubelet might delete the pods or not
      • node marked unhealthy
      • replication controllers start new pods elsewhere
  • Cluster operator error
    • Results
      • loss of pods, services, etc
      • lost of apiserver backing store
      • users unable to read API
      • etc.


  • Action: Use IaaS provider's automatic VM restarting feature for IaaS VMs

    • Mitigates: Apiserver VM shutdown or apiserver crashing
    • Mitigates: Supporting services VM shutdown or crashes
  • Action: Use IaaS providers reliable storage (e.g. GCE PD or AWS EBS volume) for VMs with apiserver+etcd

    • Mitigates: Apiserver backing storage lost
  • Action: Use high-availability configuration

    • Mitigates: Control plane node shutdown or control plane components (scheduler, API server, controller-manager) crashing
      • Will tolerate one or more simultaneous node or component failures
    • Mitigates: API server backing storage (i.e., etcd's data directory) lost
      • Assumes HA (highly-available) etcd configuration
  • Action: Snapshot apiserver PDs/EBS-volumes periodically

    • Mitigates: Apiserver backing storage lost
    • Mitigates: Some cases of operator error
    • Mitigates: Some cases of Kubernetes software fault
  • Action: use replication controller and services in front of pods

    • Mitigates: Node shutdown
    • Mitigates: Kubelet software fault
  • Action: applications (containers) designed to tolerate unexpected restarts

    • Mitigates: Node shutdown
    • Mitigates: Kubelet software fault

What's next

2.1 - Resource metrics pipeline

For Kubernetes, the Metrics API offers a basic set of metrics to support automatic scaling and similar use cases. This API makes information available about resource usage for node and pod, including metrics for CPU and memory. If you deploy the Metrics API into your cluster, clients of the Kubernetes API can then query for this information, and you can use Kubernetes' access control mechanisms to manage permissions to do so.

The HorizontalPodAutoscaler (HPA) and VerticalPodAutoscaler (VPA) use data from the metrics API to adjust workload replicas and resources to meet customer demand.

You can also view the resource metrics using the kubectl top command.

Figure 1 illustrates the architecture of the resource metrics pipeline.

flowchart RL subgraph cluster[Cluster] direction RL S[

] A[Metrics-
Server] subgraph B[Nodes] direction TB D[cAdvisor] --> C[kubelet] E[Container
runtime] --> D E1[Container
runtime] --> D P[pod data] -.- C end L[API
server] W[HPA] C ---->|Summary
API| A -->|metrics
API| L --> W end L ---> K[kubectl
top] classDef box fill:#fff,stroke:#000,stroke-width:1px,color:#000; class W,B,P,K,cluster,D,E,E1 box classDef spacewhite fill:#ffffff,stroke:#fff,stroke-width:0px,color:#000 class S spacewhite classDef k8s fill:#326ce5,stroke:#fff,stroke-width:1px,color:#fff; class A,L,C k8s

Figure 1. Resource Metrics Pipeline

The architecture components, from right to left in the figure, consist of the following:

  • cAdvisor: Daemon for collecting, aggregating and exposing container metrics included in Kubelet.

  • kubelet: Node agent for managing container resources. Resource metrics are accessible using the /metrics/resource and /stats kubelet API endpoints.

  • Summary API: API provided by the kubelet for discovering and retrieving per-node summarized stats available through the /stats endpoint.

  • metrics-server: Cluster addon component that collects and aggregates resource metrics pulled from each kubelet. The API server serves Metrics API for use by HPA, VPA, and by the kubectl top command. Metrics Server is a reference implementation of the Metrics API.

  • Metrics API: Kubernetes API supporting access to CPU and memory used for workload autoscaling. To make this work in your cluster, you need an API extension server that provides the Metrics API.

Metrics API

FEATURE STATE: Kubernetes 1.8 [beta]

The metrics-server implements the Metrics API. This API allows you to access CPU and memory usage for the nodes and pods in your cluster. Its primary role is to feed resource usage metrics to K8s autoscaler components.

Here is an example of the Metrics API request for a minikube node piped through jq for easier reading:

kubectl get --raw "/apis/" | jq '.'

Here is the same API call using curl:

curl http://localhost:8080/apis/

Sample response:

  "kind": "NodeMetrics",
  "apiVersion": "",
  "metadata": {
    "name": "minikube",
    "selfLink": "/apis/",
    "creationTimestamp": "2022-01-27T18:48:43Z"
  "timestamp": "2022-01-27T18:48:33Z",
  "window": "30s",
  "usage": {
    "cpu": "487558164n",
    "memory": "732212Ki"

Here is an example of the Metrics API request for a kube-scheduler-minikube pod contained in the kube-system namespace and piped through jq for easier reading:

kubectl get --raw "/apis/" | jq '.'

Here is the same API call using curl:

curl http://localhost:8080/apis/

Sample response:

  "kind": "PodMetrics",
  "apiVersion": "",
  "metadata": {
    "name": "kube-scheduler-minikube",
    "namespace": "kube-system",
    "selfLink": "/apis/",
    "creationTimestamp": "2022-01-27T19:25:00Z"
  "timestamp": "2022-01-27T19:24:31Z",
  "window": "30s",
  "containers": [
      "name": "kube-scheduler",
      "usage": {
        "cpu": "9559630n",
        "memory": "22244Ki"

The Metrics API is defined in the repository. You must enable the API aggregation layer and register an APIService for the API.

To learn more about the Metrics API, see resource metrics API design, the metrics-server repository and the resource metrics API.

Measuring resource usage


CPU is reported as the average core usage measured in cpu units. One cpu, in Kubernetes, is equivalent to 1 vCPU/Core for cloud providers, and 1 hyper-thread on bare-metal Intel processors.

This value is derived by taking a rate over a cumulative CPU counter provided by the kernel (in both Linux and Windows kernels). The time window used to calculate CPU is shown under window field in Metrics API.

To learn more about how Kubernetes allocates and measures CPU resources, see meaning of CPU.


Memory is reported as the working set, measured in bytes, at the instant the metric was collected.

In an ideal world, the "working set" is the amount of memory in-use that cannot be freed under memory pressure. However, calculation of the working set varies by host OS, and generally makes heavy use of heuristics to produce an estimate.

The Kubernetes model for a container's working set expects that the container runtime counts anonymous memory associated with the container in question. The working set metric typically also includes some cached (file-backed) memory, because the host OS cannot always reclaim pages.

To learn more about how Kubernetes allocates and measures memory resources, see meaning of memory.

Metrics Server

The metrics-server fetches resource metrics from the kubelets and exposes them in the Kubernetes API server through the Metrics API for use by the HPA and VPA. You can also view these metrics using the kubectl top command.

The metrics-server uses the Kubernetes API to track nodes and pods in your cluster. The metrics-server queries each node over HTTP to fetch metrics. The metrics-server also builds an internal view of pod metadata, and keeps a cache of pod health. That cached pod health information is available via the extension API that the metrics-server makes available.

For example with an HPA query, the metrics-server needs to identify which pods fulfill the label selectors in the deployment.

The metrics-server calls the kubelet API to collect metrics from each node. Depending on the metrics-server version it uses:

  • Metrics resource endpoint /metrics/resource in version v0.6.0+ or
  • Summary API endpoint /stats/summary in older versions

To learn more about the metrics-server, see the metrics-server repository.

You can also check out the following:

Summary API source

The kubelet gathers stats at the node, volume, pod and container level, and emits this information in the Summary API for consumers to read.

Here is an example of a Summary API request for a minikube node:

kubectl get --raw "/api/v1/nodes/minikube/proxy/stats/summary"

Here is the same API call using curl:

curl http://localhost:8080/api/v1/nodes/minikube/proxy/stats/summary

2.2 - Tools for Monitoring Resources

To scale an application and provide a reliable service, you need to understand how the application behaves when it is deployed. You can examine application performance in a Kubernetes cluster by examining the containers, pods, services, and the characteristics of the overall cluster. Kubernetes provides detailed information about an application's resource usage at each of these levels. This information allows you to evaluate your application's performance and where bottlenecks can be removed to improve overall performance.

In Kubernetes, application monitoring does not depend on a single monitoring solution. On new clusters, you can use resource metrics or full metrics pipelines to collect monitoring statistics.

Resource metrics pipeline

The resource metrics pipeline provides a limited set of metrics related to cluster components such as the Horizontal Pod Autoscaler controller, as well as the kubectl top utility. These metrics are collected by the lightweight, short-term, in-memory metrics-server and are exposed via the API.

metrics-server discovers all nodes on the cluster and queries each node's kubelet for CPU and memory usage. The kubelet acts as a bridge between the Kubernetes master and the nodes, managing the pods and containers running on a machine. The kubelet translates each pod into its constituent containers and fetches individual container usage statistics from the container runtime through the container runtime interface. If you use a container runtime that uses Linux cgroups and namespaces to implement containers, and the container runtime does not publish usage statistics, then the kubelet can look up those statistics directly (using code from cAdvisor). No matter how those statistics arrive, the kubelet then exposes the aggregated pod resource usage statistics through the metrics-server Resource Metrics API. This API is served at /metrics/resource/v1beta1 on the kubelet's authenticated and read-only ports.

Full metrics pipeline

A full metrics pipeline gives you access to richer metrics. Kubernetes can respond to these metrics by automatically scaling or adapting the cluster based on its current state, using mechanisms such as the Horizontal Pod Autoscaler. The monitoring pipeline fetches metrics from the kubelet and then exposes them to Kubernetes via an adapter by implementing either the or API.

Prometheus, a CNCF project, can natively monitor Kubernetes, nodes, and Prometheus itself. Full metrics pipeline projects that are not part of the CNCF are outside the scope of Kubernetes documentation.

What's next

Learn about additional debugging tools, including:

2.3 - Monitor Node Health

Node Problem Detector is a daemon for monitoring and reporting about a node's health. You can run Node Problem Detector as a DaemonSet or as a standalone daemon. Node Problem Detector collects information about node problems from various daemons and reports these conditions to the API server as NodeCondition and Event.

To learn how to install and use Node Problem Detector, see Node Problem Detector project documentation.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:


  • Node Problem Detector only supports file based kernel log. Log tools such as journald are not supported.

  • Node Problem Detector uses the kernel log format for reporting kernel issues. To learn how to extend the kernel log format, see Add support for another log format.

Enabling Node Problem Detector

Some cloud providers enable Node Problem Detector as an Addon. You can also enable Node Problem Detector with kubectl or by creating an Addon pod.

Using kubectl to enable Node Problem Detector

kubectl provides the most flexible management of Node Problem Detector. You can overwrite the default configuration to fit it into your environment or to detect customized node problems. For example:

  1. Create a Node Problem Detector configuration similar to node-problem-detector.yaml:

    apiVersion: apps/v1
       kind: DaemonSet
         name: node-problem-detector-v0.1
         namespace: kube-system
           k8s-app: node-problem-detector
           version: v0.1
             k8s-app: node-problem-detector  
             version: v0.1
               k8s-app: node-problem-detector
               version: v0.1
             hostNetwork: true
             - name: node-problem-detector
                 privileged: true
                   cpu: "200m"
                   memory: "100Mi"
                   cpu: "20m"
                   memory: "20Mi"
               - name: log
                 mountPath: /log
                 readOnly: true
             - name: log
                 path: /var/log/
  2. Start node problem detector with kubectl:

    kubectl apply -f

Using an Addon pod to enable Node Problem Detector

If you are using a custom cluster bootstrap solution and don't need to overwrite the default configuration, you can leverage the Addon pod to further automate the deployment.

Create node-problem-detector.yaml, and save the configuration in the Addon pod's directory /etc/kubernetes/addons/node-problem-detector on a control plane node.

Overwrite the configuration

The default configuration is embedded when building the Docker image of Node Problem Detector.

However, you can use a ConfigMap to overwrite the configuration:

  1. Change the configuration files in config/

  2. Create the ConfigMap node-problem-detector-config:

    kubectl create configmap node-problem-detector-config --from-file=config/
  3. Change the node-problem-detector.yaml to use the ConfigMap:

    apiVersion: apps/v1
       kind: DaemonSet
         name: node-problem-detector-v0.1
         namespace: kube-system
           k8s-app: node-problem-detector
           version: v0.1
             k8s-app: node-problem-detector  
             version: v0.1
               k8s-app: node-problem-detector
               version: v0.1
             hostNetwork: true
             - name: node-problem-detector
                 privileged: true
                   cpu: "200m"
                   memory: "100Mi"
                   cpu: "20m"
                   memory: "20Mi"
               - name: log
                 mountPath: /log
                 readOnly: true
               - name: config # Overwrite the config/ directory with ConfigMap volume
                 mountPath: /config
                 readOnly: true
             - name: log
                 path: /var/log/
             - name: config # Define ConfigMap volume
                 name: node-problem-detector-config
  4. Recreate the Node Problem Detector with the new configuration file:

    # If you have a node-problem-detector running, delete before recreating
    kubectl delete -f
    kubectl apply -f

Overwriting a configuration is not supported if a Node Problem Detector runs as a cluster Addon. The Addon manager does not support ConfigMap.

Kernel Monitor

Kernel Monitor is a system log monitor daemon supported in the Node Problem Detector. Kernel monitor watches the kernel log and detects known kernel issues following predefined rules.

The Kernel Monitor matches kernel issues according to a set of predefined rule list in config/kernel-monitor.json. The rule list is extensible. You can expand the rule list by overwriting the configuration.

Add new NodeConditions

To support a new NodeCondition, create a condition definition within the conditions field in config/kernel-monitor.json, for example:

  "type": "NodeConditionType",
  "reason": "CamelCaseDefaultNodeConditionReason",
  "message": "arbitrary default node condition message"

Detect new problems

To detect new problems, you can extend the rules field in config/kernel-monitor.json with a new rule definition:

  "type": "temporary/permanent",
  "condition": "NodeConditionOfPermanentIssue",
  "reason": "CamelCaseShortReason",
  "message": "regexp matching the issue in the kernel log"

Configure path for the kernel log device

Check your kernel log path location in your operating system (OS) distribution. The Linux kernel log device is usually presented as /dev/kmsg. However, the log path location varies by OS distribution. The log field in config/kernel-monitor.json represents the log path inside the container. You can configure the log field to match the device path as seen by the Node Problem Detector.

Add support for another log format

Kernel monitor uses the Translator plugin to translate the internal data structure of the kernel log. You can implement a new translator for a new log format.

Recommendations and restrictions

It is recommended to run the Node Problem Detector in your cluster to monitor node health. When running the Node Problem Detector, you can expect extra resource overhead on each node. Usually this is fine, because:

  • The kernel log grows relatively slowly.
  • A resource limit is set for the Node Problem Detector.
  • Even under high load, the resource usage is acceptable. For more information, see the Node Problem Detector benchmark result.

2.4 - Debugging Kubernetes nodes with crictl

FEATURE STATE: Kubernetes v1.11 [stable]

crictl is a command-line interface for CRI-compatible container runtimes. You can use it to inspect and debug container runtimes and applications on a Kubernetes node. crictl and its source are hosted in the cri-tools repository.

Before you begin

crictl requires a Linux operating system with a CRI runtime.

Installing crictl

You can download a compressed archive crictl from the cri-tools release page, for several different architectures. Download the version that corresponds to your version of Kubernetes. Extract it and move it to a location on your system path, such as /usr/local/bin/.

General usage

The crictl command has several subcommands and runtime flags. Use crictl help or crictl <subcommand> help for more details.

You can set the endpoint for crictl by doing one of the following:

  • Set the --runtime-endpoint and --image-endpoint flags.
  • Set the endpoint in the configuration file /etc/crictl.yaml. To specify a different file, use the --config=PATH_TO_FILE flag when you run crictl.

You can also specify timeout values when connecting to the server and enable or disable debugging, by specifying timeout or debug values in the configuration file or using the --timeout and --debug command-line flags.

To view or edit the current configuration, view or edit the contents of /etc/crictl.yaml. For example, the configuration when using the containerd container runtime would be similar to this:

runtime-endpoint: unix:///var/run/containerd/containerd.sock
image-endpoint: unix:///var/run/containerd/containerd.sock
timeout: 10
debug: true

To learn more about crictl, refer to the crictl documentation.

Example crictl commands

The following examples show some crictl commands and example output.

List pods

List all pods:

crictl pods

The output is similar to this:

POD ID              CREATED              STATE               NAME                         NAMESPACE           ATTEMPT
926f1b5a1d33a       About a minute ago   Ready               sh-84d7dcf559-4r2gq          default             0
4dccb216c4adb       About a minute ago   Ready               nginx-65899c769f-wv2gp       default             0
a86316e96fa89       17 hours ago         Ready               kube-proxy-gblk4             kube-system         0
919630b8f81f1       17 hours ago         Ready               nvidia-device-plugin-zgbbv   kube-system         0

List pods by name:

crictl pods --name nginx-65899c769f-wv2gp

The output is similar to this:

POD ID              CREATED             STATE               NAME                     NAMESPACE           ATTEMPT
4dccb216c4adb       2 minutes ago       Ready               nginx-65899c769f-wv2gp   default             0

List pods by label:

crictl pods --label run=nginx

The output is similar to this:

POD ID              CREATED             STATE               NAME                     NAMESPACE           ATTEMPT
4dccb216c4adb       2 minutes ago       Ready               nginx-65899c769f-wv2gp   default             0

List images

List all images:

crictl images

The output is similar to this:

IMAGE                                     TAG                 IMAGE ID            SIZE
busybox                                   latest              8c811b4aec35f       1.15MB   v1.10.3             e179bbfe5d238       665MB       3.1                 da86e6ba6ca19       742kB
nginx                                     latest              cd5239a0906a6       109MB

List images by repository:

crictl images nginx

The output is similar to this:

IMAGE               TAG                 IMAGE ID            SIZE
nginx               latest              cd5239a0906a6       109MB

Only list image IDs:

crictl images -q

The output is similar to this:


List containers

List all containers:

crictl ps -a

The output is similar to this:

CONTAINER ID        IMAGE                                                                                                             CREATED             STATE               NAME                       ATTEMPT
1f73f2d81bf98       busybox@sha256:141c253bc4c3fd0a201d32dc1f493bcf3fff003b6df416dea4f41046e0f37d47                                   7 minutes ago       Running             sh                         1
9c5951df22c78       busybox@sha256:141c253bc4c3fd0a201d32dc1f493bcf3fff003b6df416dea4f41046e0f37d47                                   8 minutes ago       Exited              sh                         0
87d3992f84f74       nginx@sha256:d0a8828cccb73397acb0073bf34f4d7d8aa315263f1e7806bf8c55d8ac139d5f                                     8 minutes ago       Running             nginx                      0
1941fb4da154f   18 hours ago        Running             kube-proxy                 0

List running containers:

crictl ps

The output is similar to this:

CONTAINER ID        IMAGE                                                                                                             CREATED             STATE               NAME                       ATTEMPT
1f73f2d81bf98       busybox@sha256:141c253bc4c3fd0a201d32dc1f493bcf3fff003b6df416dea4f41046e0f37d47                                   6 minutes ago       Running             sh                         1
87d3992f84f74       nginx@sha256:d0a8828cccb73397acb0073bf34f4d7d8aa315263f1e7806bf8c55d8ac139d5f                                     7 minutes ago       Running             nginx                      0
1941fb4da154f   17 hours ago        Running             kube-proxy                 0

Execute a command in a running container

crictl exec -i -t 1f73f2d81bf98 ls

The output is similar to this:

bin   dev   etc   home  proc  root  sys   tmp   usr   var

Get a container's logs

Get all container logs:

crictl logs 87d3992f84f74

The output is similar to this: - - [06/Jun/2018:02:45:49 +0000] "GET / HTTP/1.1" 200 612 "-" "curl/7.47.0" "-" - - [06/Jun/2018:02:45:50 +0000] "GET / HTTP/1.1" 200 612 "-" "curl/7.47.0" "-" - - [06/Jun/2018:02:45:51 +0000] "GET / HTTP/1.1" 200 612 "-" "curl/7.47.0" "-"

Get only the latest N lines of logs:

crictl logs --tail=1 87d3992f84f74

The output is similar to this: - - [06/Jun/2018:02:45:51 +0000] "GET / HTTP/1.1" 200 612 "-" "curl/7.47.0" "-"

Run a pod sandbox

Using crictl to run a pod sandbox is useful for debugging container runtimes. On a running Kubernetes cluster, the sandbox will eventually be stopped and deleted by the Kubelet.

  1. Create a JSON file like the following:

      "metadata": {
        "name": "nginx-sandbox",
        "namespace": "default",
        "attempt": 1,
        "uid": "hdishd83djaidwnduwk28bcsb"
      "logDirectory": "/tmp",
      "linux": {
  2. Use the crictl runp command to apply the JSON and run the sandbox.

    crictl runp pod-config.json

    The ID of the sandbox is returned.

Create a container

Using crictl to create a container is useful for debugging container runtimes. On a running Kubernetes cluster, the sandbox will eventually be stopped and deleted by the Kubelet.

  1. Pull a busybox image

    crictl pull busybox
    Image is up to date for busybox@sha256:141c253bc4c3fd0a201d32dc1f493bcf3fff003b6df416dea4f41046e0f37d47
  2. Create configs for the pod and the container:

    Pod config:

      "metadata": {
        "name": "nginx-sandbox",
        "namespace": "default",
        "attempt": 1,
        "uid": "hdishd83djaidwnduwk28bcsb"
      "log_directory": "/tmp",
      "linux": {

    Container config:

      "metadata": {
        "name": "busybox"
        "image": "busybox"
      "command": [
      "linux": {
  3. Create the container, passing the ID of the previously-created pod, the container config file, and the pod config file. The ID of the container is returned.

    crictl create f84dd361f8dc51518ed291fbadd6db537b0496536c1d2d6c05ff943ce8c9a54f container-config.json pod-config.json
  4. List all containers and verify that the newly-created container has its state set to Created.

    crictl ps -a

    The output is similar to this:

    CONTAINER ID        IMAGE               CREATED             STATE               NAME                ATTEMPT
    3e025dd50a72d       busybox             32 seconds ago      Created             busybox             0

Start a container

To start a container, pass its ID to crictl start:

crictl start 3e025dd50a72d956c4f14881fbb5b1080c9275674e95fb67f965f6478a957d60

The output is similar to this:


Check the container has its state set to Running.

crictl ps

The output is similar to this:

3e025dd50a72d  busybox  About a minute ago   Running  busybox  0

What's next

2.5 - Auditing

Kubernetes auditing provides a security-relevant, chronological set of records documenting the sequence of actions in a cluster. The cluster audits the activities generated by users, by applications that use the Kubernetes API, and by the control plane itself.

Auditing allows cluster administrators to answer the following questions:

  • what happened?
  • when did it happen?
  • who initiated it?
  • on what did it happen?
  • where was it observed?
  • from where was it initiated?
  • to where was it going?

Audit records begin their lifecycle inside the kube-apiserver component. Each request on each stage of its execution generates an audit event, which is then pre-processed according to a certain policy and written to a backend. The policy determines what's recorded and the backends persist the records. The current backend implementations include logs files and webhooks.

Each request can be recorded with an associated stage. The defined stages are:

  • RequestReceived - The stage for events generated as soon as the audit handler receives the request, and before it is delegated down the handler chain.
  • ResponseStarted - Once the response headers are sent, but before the response body is sent. This stage is only generated for long-running requests (e.g. watch).
  • ResponseComplete - The response body has been completed and no more bytes will be sent.
  • Panic - Events generated when a panic occurred.

The audit logging feature increases the memory consumption of the API server because some context required for auditing is stored for each request. Memory consumption depends on the audit logging configuration.

Audit policy

Audit policy defines rules about what events should be recorded and what data they should include. The audit policy object structure is defined in the API group. When an event is processed, it's compared against the list of rules in order. The first matching rule sets the audit level of the event. The defined audit levels are:

  • None - don't log events that match this rule.
  • Metadata - log request metadata (requesting user, timestamp, resource, verb, etc.) but not request or response body.
  • Request - log event metadata and request body but not response body. This does not apply for non-resource requests.
  • RequestResponse - log event metadata, request and response bodies. This does not apply for non-resource requests.

You can pass a file with the policy to kube-apiserver using the --audit-policy-file flag. If the flag is omitted, no events are logged. Note that the rules field must be provided in the audit policy file. A policy with no (0) rules is treated as illegal.

Below is an example audit policy file:

apiVersion: # This is required.
kind: Policy
# Don't generate audit events for all requests in RequestReceived stage.
  - "RequestReceived"
  # Log pod changes at RequestResponse level
  - level: RequestResponse
    - group: ""
      # Resource "pods" doesn't match requests to any subresource of pods,
      # which is consistent with the RBAC policy.
      resources: ["pods"]
  # Log "pods/log", "pods/status" at Metadata level
  - level: Metadata
    - group: ""
      resources: ["pods/log", "pods/status"]

  # Don't log requests to a configmap called "controller-leader"
  - level: None
    - group: ""
      resources: ["configmaps"]
      resourceNames: ["controller-leader"]

  # Don't log watch requests by the "system:kube-proxy" on endpoints or services
  - level: None
    users: ["system:kube-proxy"]
    verbs: ["watch"]
    - group: "" # core API group
      resources: ["endpoints", "services"]

  # Don't log authenticated requests to certain non-resource URL paths.
  - level: None
    userGroups: ["system:authenticated"]
    - "/api*" # Wildcard matching.
    - "/version"

  # Log the request body of configmap changes in kube-system.
  - level: Request
    - group: "" # core API group
      resources: ["configmaps"]
    # This rule only applies to resources in the "kube-system" namespace.
    # The empty string "" can be used to select non-namespaced resources.
    namespaces: ["kube-system"]

  # Log configmap and secret changes in all other namespaces at the Metadata level.
  - level: Metadata
    - group: "" # core API group
      resources: ["secrets", "configmaps"]

  # Log all other resources in core and extensions at the Request level.
  - level: Request
    - group: "" # core API group
    - group: "extensions" # Version of group should NOT be included.

  # A catch-all rule to log all other requests at the Metadata level.
  - level: Metadata
    # Long-running requests like watches that fall under this rule will not
    # generate an audit event in RequestReceived.
      - "RequestReceived"

You can use a minimal audit policy file to log all requests at the Metadata level:

# Log all requests at the Metadata level.
kind: Policy
- level: Metadata

If you're crafting your own audit profile, you can use the audit profile for Google Container-Optimized OS as a starting point. You can check the script, which generates an audit policy file. You can see most of the audit policy file by looking directly at the script.

You can also refer to the Policy configuration reference for details about the fields defined.

Audit backends

Audit backends persist audit events to an external storage. Out of the box, the kube-apiserver provides two backends:

  • Log backend, which writes events into the filesystem
  • Webhook backend, which sends events to an external HTTP API

In all cases, audit events follow a structure defined by the Kubernetes API in the API group.

Log backend

The log backend writes audit events to a file in JSONlines format. You can configure the log audit backend using the following kube-apiserver flags:

  • --audit-log-path specifies the log file path that log backend uses to write audit events. Not specifying this flag disables log backend. - means standard out
  • --audit-log-maxage defined the maximum number of days to retain old audit log files
  • --audit-log-maxbackup defines the maximum number of audit log files to retain
  • --audit-log-maxsize defines the maximum size in megabytes of the audit log file before it gets rotated

If your cluster's control plane runs the kube-apiserver as a Pod, remember to mount the hostPath to the location of the policy file and log file, so that audit records are persisted. For example:

    --audit-policy-file=/etc/kubernetes/audit-policy.yaml \

then mount the volumes:

  - mountPath: /etc/kubernetes/audit-policy.yaml
    name: audit
    readOnly: true
  - mountPath: /var/log/kubernetes/audit/
    name: audit-log
    readOnly: false

and finally configure the hostPath:

- name: audit
    path: /etc/kubernetes/audit-policy.yaml
    type: File

- name: audit-log
    path: /var/log/kubernetes/audit/
    type: DirectoryOrCreate

Webhook backend

The webhook audit backend sends audit events to a remote web API, which is assumed to be a form of the Kubernetes API, including means of authentication. You can configure a webhook audit backend using the following kube-apiserver flags:

  • --audit-webhook-config-file specifies the path to a file with a webhook configuration. The webhook configuration is effectively a specialized kubeconfig.
  • --audit-webhook-initial-backoff specifies the amount of time to wait after the first failed request before retrying. Subsequent requests are retried with exponential backoff.

The webhook config file uses the kubeconfig format to specify the remote address of the service and credentials used to connect to it.

Event batching

Both log and webhook backends support batching. Using webhook as an example, here's the list of available flags. To get the same flag for log backend, replace webhook with log in the flag name. By default, batching is enabled in webhook and disabled in log. Similarly, by default throttling is enabled in webhook and disabled in log.

  • --audit-webhook-mode defines the buffering strategy. One of the following:
    • batch - buffer events and asynchronously process them in batches. This is the default.
    • blocking - block API server responses on processing each individual event.
    • blocking-strict - Same as blocking, but when there is a failure during audit logging at the RequestReceived stage, the whole request to the kube-apiserver fails.

The following flags are used only in the batch mode:

  • --audit-webhook-batch-buffer-size defines the number of events to buffer before batching. If the rate of incoming events overflows the buffer, events are dropped.
  • --audit-webhook-batch-max-size defines the maximum number of events in one batch.
  • --audit-webhook-batch-max-wait defines the maximum amount of time to wait before unconditionally batching events in the queue.
  • --audit-webhook-batch-throttle-qps defines the maximum average number of batches generated per second.
  • --audit-webhook-batch-throttle-burst defines the maximum number of batches generated at the same moment if the allowed QPS was underutilized previously.

Parameter tuning

Parameters should be set to accommodate the load on the API server.

For example, if kube-apiserver receives 100 requests each second, and each request is audited only on ResponseStarted and ResponseComplete stages, you should account for ≅200 audit events being generated each second. Assuming that there are up to 100 events in a batch, you should set throttling level at least 2 queries per second. Assuming that the backend can take up to 5 seconds to write events, you should set the buffer size to hold up to 5 seconds of events; that is: 10 batches, or 1000 events.

In most cases however, the default parameters should be sufficient and you don't have to worry about setting them manually. You can look at the following Prometheus metrics exposed by kube-apiserver and in the logs to monitor the state of the auditing subsystem.

  • apiserver_audit_event_total metric contains the total number of audit events exported.
  • apiserver_audit_error_total metric contains the total number of events dropped due to an error during exporting.

Log entry truncation

Both log and webhook backends support limiting the size of events that are logged. As an example, the following is the list of flags available for the log backend:

  • audit-log-truncate-enabled whether event and batch truncating is enabled.
  • audit-log-truncate-max-batch-size maximum size in bytes of the batch sent to the underlying backend.
  • audit-log-truncate-max-event-size maximum size in bytes of the audit event sent to the underlying backend.

By default truncate is disabled in both webhook and log, a cluster administrator should set audit-log-truncate-enabled or audit-webhook-truncate-enabled to enable the feature.

What's next

2.6 - Developing and debugging services locally using telepresence

Kubernetes applications usually consist of multiple, separate services, each running in its own container. Developing and debugging these services on a remote Kubernetes cluster can be cumbersome, requiring you to get a shell on a running container in order to run debugging tools.

telepresence is a tool to ease the process of developing and debugging services locally while proxying the service to a remote Kubernetes cluster. Using telepresence allows you to use custom tools, such as a debugger and IDE, for a local service and provides the service full access to ConfigMap, secrets, and the services running on the remote cluster.

This document describes using telepresence to develop and debug services running on a remote cluster locally.

Before you begin

  • Kubernetes cluster is installed
  • kubectl is configured to communicate with the cluster
  • Telepresence is installed

Connecting your local machine to a remote Kubernetes cluster

After installing telepresence, run telepresence connect to launch its Daemon and connect your local workstation to the cluster.

$ telepresence connect
Launching Telepresence Daemon
Connected to context default (https://<cluster public IP>)

You can curl services using the Kubernetes syntax e.g. curl -ik https://kubernetes.default

Developing or debugging an existing service

When developing an application on Kubernetes, you typically program or debug a single service. The service might require access to other services for testing and debugging. One option is to use the continuous deployment pipeline, but even the fastest deployment pipeline introduces a delay in the program or debug cycle.

Use the telepresence intercept $SERVICE_NAME --port $LOCAL_PORT:REMOTE_PORT command to create an "intercept" for rerouting remote service traffic.


  • $SERVICE_NAME is the name of your local service
  • $LOCAL_PORT is the port that your service is running on your local workstation
  • And $REMOTE_PORT is the port your service listens to in the cluster

Running this command tells Telepresence to send remote traffic to your local service instead of the service in the remote Kubernetes cluster. Make edits to your service source code locally, save, and see the corresponding changes when accessing your remote application take effect immediately. You can also run your local service using a debugger or any other local development tool.

How does Telepresence work?

Telepresence installs a traffic-agent sidecar next to your existing application's container running in the remote cluster. It then captures all traffic requests going into the Pod, and instead of forwarding this to the application in the remote cluster, it routes all traffic (when you create a global intercept) or a subset of the traffic (when you create a personal intercept) to your local development environment.

What's next

If you're interested in a hands-on tutorial, check out this tutorial that walks through locally developing the Guestbook application on Google Kubernetes Engine.

For further reading, visit the Telepresence website.

2.7 - Windows debugging tips

Node-level troubleshooting

  1. My Pods are stuck at "Container Creating" or restarting over and over

    Ensure that your pause image is compatible with your Windows OS version. See Pause container to see the latest / recommended pause image and/or get more information.

  2. My pods show status as ErrImgPull or ImagePullBackOff

    Ensure that your Pod is getting scheduled to a compatible Windows Node.

    More information on how to specify a compatible node for your Pod can be found in this guide.

Network troubleshooting

  1. My Windows Pods do not have network connectivity

    If you are using virtual machines, ensure that MAC spoofing is enabled on all the VM network adapter(s).

  2. My Windows Pods cannot ping external resources

    Windows Pods do not have outbound rules programmed for the ICMP protocol. However, TCP/UDP is supported. When trying to demonstrate connectivity to resources outside of the cluster, substitute ping <IP> with corresponding curl <IP> commands.

    If you are still facing problems, most likely your network configuration in cni.conf deserves some extra attention. You can always edit this static file. The configuration update will apply to any new Kubernetes resources.

    One of the Kubernetes networking requirements (see Kubernetes model) is for cluster communication to occur without NAT internally. To honor this requirement, there is an ExceptionList for all the communication where you do not want outbound NAT to occur. However, this also means that you need to exclude the external IP you are trying to query from the ExceptionList. Only then will the traffic originating from your Windows pods be SNAT'ed correctly to receive a response from the outside world. In this regard, your ExceptionList in cni.conf should look as follows:

    "ExceptionList": [
                    "",  # Cluster subnet
                    "",   # Service subnet
                    "" # Management (host) subnet
  3. My Windows node cannot access NodePort type Services

    Local NodePort access from the node itself fails. This is a known limitation. NodePort access works from other nodes or external clients.

  4. vNICs and HNS endpoints of containers are being deleted

    This issue can be caused when the hostname-override parameter is not passed to kube-proxy. To resolve it, users need to pass the hostname to kube-proxy as follows:

    C:\k\kube-proxy.exe --hostname-override=$(hostname)
  5. My Windows node cannot access my services using the service IP

    This is a known limitation of the networking stack on Windows. However, Windows Pods can access the Service IP.

  6. No network adapter is found when starting the kubelet

    The Windows networking stack needs a virtual adapter for Kubernetes networking to work. If the following commands return no results (in an admin shell), virtual network creation — a necessary prerequisite for the kubelet to work — has failed:

    Get-HnsNetwork | ? Name -ieq "cbr0"
    Get-NetAdapter | ? Name -Like "vEthernet (Ethernet*"

    Often it is worthwhile to modify the InterfaceName parameter of the start.ps1 script, in cases where the host's network adapter isn't "Ethernet". Otherwise, consult the output of the start-kubelet.ps1 script to see if there are errors during virtual network creation.

  7. DNS resolution is not properly working

    Check the DNS limitations for Windows in this section.

  8. kubectl port-forward fails with "unable to do port forwarding: wincat not found"

    This was implemented in Kubernetes 1.15 by including wincat.exe in the pause infrastructure container Be sure to use a supported version of Kubernetes. If you would like to build your own pause infrastructure container be sure to include wincat.

  9. My Kubernetes installation is failing because my Windows Server node is behind a proxy

    If you are behind a proxy, the following PowerShell environment variables must be defined:

    [Environment]::SetEnvironmentVariable("HTTP_PROXY", "", [EnvironmentVariableTarget]::Machine)
    [Environment]::SetEnvironmentVariable("HTTPS_PROXY", "", [EnvironmentVariableTarget]::Machine)

Flannel troubleshooting

  1. With Flannel, my nodes are having issues after rejoining a cluster

    Whenever a previously deleted node is being re-joined to the cluster, flannelD tries to assign a new pod subnet to the node. Users should remove the old pod subnet configuration files in the following paths:

    Remove-Item C:\k\SourceVip.json
    Remove-Item C:\k\SourceVipRequest.json
  2. Flanneld is stuck in "Waiting for the Network to be created"

    There are numerous reports of this issue; most likely it is a timing issue for when the management IP of the flannel network is set. A workaround is to relaunch start.ps1 or relaunch it manually as follows:

    [Environment]::SetEnvironmentVariable("NODE_NAME", "<Windows_Worker_Hostname>")
    C:\flannel\flanneld.exe --kubeconfig-file=c:\k\config --iface=<Windows_Worker_Node_IP> --ip-masq=1 --kube-subnet-mgr=1
  3. My Windows Pods cannot launch because of missing /run/flannel/subnet.env

    This indicates that Flannel didn't launch correctly. You can either try to restart flanneld.exe or you can copy the files over manually from /run/flannel/subnet.env on the Kubernetes master to C:\run\flannel\subnet.env on the Windows worker node and modify the FLANNEL_SUBNET row to a different number. For example, if node subnet is desired:


Further investigation

If these steps don't resolve your problem, you can get help running Windows containers on Windows nodes in Kubernetes through: