[00:00] Hi everyone today we will cover cluster architecture for CKA domain 2 troubleshooting and guide you through the cluster and ARTME the processes that run in it and exactly where to look when things go wrong. Let's dive in.
[00:16] This video is aimed at CKA candidates if you are studying for the exam treat this as a focus troubleshooting checklist. Understand what each component does, how they talk to each
[00:33] other and the practical commands and check to run when something breaks. In domain 2, we focus on 4 areas, the cluster and its nodes, the core cluster components,
[00:48] monitoring and resources, check and services and networking. I'll walk you through each area in sequence so you know both the theory and the troubleshooting steps.
[01:02] At the highest level a Kubernetes cluster is a group of machines working together, a control plane that makes decisions and worker nodes that runs your workloads.
[01:18] If you are looking at the diagram on the slide, imagine the control plane at the top and worker and worker nodes below it. The control plane holds components that accept request and record
[01:30] state. Worker nodes run pods and actually execute the containers. Troubleshooting starts by asking is the problem with the decision making which is the control plane or the execution
[01:43] which is the worker node. Now first the node agent which is the cubelet. It runs on every node. It continuously reports node status to the control plane. Its core job is managing
[01:58] container on its node, pulling images, starting and stopping container via the container runtime. Enforcing the pod specs the API server provides. When the API server gives an instruction
[02:16] like create, update or delete a pod, Kubelet implements it. Kubelet also collects resource usage on the node and invokes network and storage plugins
[02:28] so pods get their networking and volumes. If you need to prob the Kubelet, check its pods and expose antipoints for metrics and pod information. these are the places to look when pods fail to start or nodes report the old conditions
[02:46] now coming towards the demo in this demo we will disable the unauthorized and authenticated access via Kubelet
[02:59] now first checking the services of the Kubelet Here you can see the configuration file location.
[03:16] Now accessing this file file as you can see that anonymous authentication is enabled
[03:33] the read only port is also available. Now first taking the normal operation if I get the ports command to direct link or a direct request you can see all the pod information
[03:50] because the request is not authenticated. Now first changing the anonymous authentication changing it to true changing the true to false and then authorization we change it to webhook
[04:12] and for read only ports we don't want that anyone access the matrix as well so we configure 0 as the 0 to only port now restarting the kubelet
[04:31] as you can see now we are not able to access the pods through direct requests and we are not able to see the matrix as well so in this form you can enable the
[04:44] kubelet security or you can check for any nif configuration in the kubelet configuration file that is obtained through the through checking the services of the kubelet.
[04:58] Now moving forward kubelet talks to the container runtime through a standard interface called the CRI. This makes kubelet runtime agnostic.
[05:10] Because of the CRI which is container runtime interface, Kubernetes works with container D, CRIO or other runtime without changing the core logic. Container D is a common lightweight runtime you will see in modern clusters.
[05:24] For troubleshooting, containers directly use the runtime-aware CLI, for example CRI-CTL or normally we call it CHI-CTL. to list the containers, inspect them, check the logs and execute the into the running containers.
[05:41] This helps you separate container level problems from the Kubernetes level problem. In the next demo we will be using the KICTL commands to check the logs and check the status
[05:56] of the different containers that are running on the control plane. Now moving forward, Q-Proxy runs on each node and implements the networking rules that let the services reach the correct pods.
[06:10] Now think of it as the traffic director. When a service is created, Q-Proxy sets up the local networking rules to forward traffic to the correct pod endpoints. In many clusters, Q-Proxy runs as a pod in the control plan and namespace and is deployed
[06:26] as a daemon set so every node has one and when you delete it it will automatically be created if client can't reach a service check the kube proxy logs and the local forwarding rules
[06:40] on the node that's often where services reachability problem shows up now moving forward we come towards the kube api server the api server is the cluster's front door
[06:56] Everything like kubectl commands, controller, kubelets talk to it. It authenticates, authorizes, validates, requests, run admission controllers and persists cluster state to the cluster store.
[07:11] When you create an object the API server is responsible for receiving that request, first validating it and then the scoring the status changes if any are requested. If the cluster is unresponsive or kubectl calls fails, the API server is the first point to check.
[07:31] Now for quick troubleshooting, what happens is like the client request arrives at the API server. The API server first validates, authenticates and persists the object.
[07:43] Other components like scheduler, controller watch for the changes and react accordingly. If something stops at any step the API server logs and the static manifest on the control plane nodes are your first checkpoint.
[07:57] Now coming towards live demo In this live demo we will be troubleshooting the kube API As you can see when I run the kubectl command it is unresponsive and the error shows that 6443 was refused
[08:16] So 6443 is the port for the kube API. Now checking the cryctl first and checking the container there is no container being
[08:28] built. Now checking the log files there is no log file created as well. So the actually the container has not run for once as well.
[08:48] Now checking the logs. Now moving towards the configuration file. The manifest for the in the cube API.
[09:02] Now you see that the metadata there is a semicolon using sudo for colon. Now if we change it this is the one error.
[09:14] Now let's see if the if we can run the kubectl command again. Let's first check the container. Now the container has been built.
[09:27] The container is available. we can check the logs of this container
[09:39] now if you see the logs again checking the container giving the proper container ID now you can see that there is an unknown flag and instead of mob
[10:01] modus has been written now we can check on the configuration file and make the correction accordingly
[10:19] now again checking if the container is now available eqdapi container through crictl command now checking the logs again
[10:40] now if you see that there is a connection refusal and a port number 23000 has been used now checking the 23000 port now you can see that the etcd is configured with 23000 port
[10:56] configured with 23000 ports but the actual port number is 2379
[11:08] now after making the changes lets see if the kube api server is now responding and we are able to run the kubectl command again
[11:20] Yes, now it is working. So there were three misconfigurations and we were able to find those misconfigurations through the different kinds of logs that are available.
[11:38] Now moving forward the controller manager runs controllers that ensures the clusters actual state matches with the desired state. like replica sets, jobs, endpoints, procession volume, service accounts and more.
[11:52] Controllers continuously compare current state to the desired state and make the changes to converge that state. For example, restarting pods when replicas drop.
[12:04] Controllers don receive external requests directly they act through API server If the resources aren reconciling as expected the control manager logs in it watch the cycle and its watch cycle are there
[12:19] to investigate. Now coming towards the live demo.
[12:32] Now as we can see that all these control components are present in the cube system name space and here the controller is not working.
[12:50] Now checking the logs of the particular pod.
[13:05] In the logs we can see that there is an unknown flag that is sidecar insertion.
[13:17] Now let us open the configuration file of kube controller and find the flag that is causing the error.
[13:32] Now removing the line which is causing the problem and then again see whether the pod is not available is now available or not as you can see now the pod is running
[14:01] now moving forward etc is the classic persistent store the single source of truth it stores all the objects in configuration like pause, nodes, secrets, roles, everything.
[14:17] Any change you make is first written to ETCD. Other components read or watch ETCD for updates. If ETCD becomes unhealthy, the control plane loses its memory that can break the scheduling,
[14:32] reconciliation and more. For diagnosis, confirm cluster membership, health and death rights are succeeding.
[14:44] Now moving forward the Cubed Scheduler. The Scheduler is responsible for choosing the best node for each pod. It filters out nodes that don't meet requirement, then ranks the remaining nodes and picks the
[14:58] best match based on the scoring rules. It watches the API server for unscheduled pods then assign them. If pods are stuck in the pending state, scheduler condition and logs are the right place to
[15:12] look for. Check why pods fail filters or why the ranking pick a node that later rejected the pod. So if the pod is not scheduling, the control panel you want to check is the Qt scheduler.
[15:29] Now let us wrap it up. During troubleshooting, determine whether issue exists at control plan or the node level.
[15:41] The first step is to determine where to look for the control plan or the node level. Now the control plan has multiple control components. Then we have to follow the components specific to each problem.
[15:56] See their configuration file, know the configuration file location and the specific ports and try to interpret the logs collected from different sources.
[16:09] If this helped, like and subscribe the channel and good luck for