Conf42 Site Reliability Engineering (SRE) 2024 - Online

2024: I Don't Know K8S and at This Point, I'm Too Afraid To Ask

Video size:


It’s 2024 and everybody has used or at least heard about Kubernetes… If you’re at the start of the journey into Kubernetes, or you’d like to understand why Kubernetes is a compute abstraction and a platform for building platforms, then this talk is for you.


  • Today's presentation is entitled I don't know kubernetes and at this point I'm too afraid to ask. We're gonna touch upon Kubernetes API, Kubers controller and operator pattern. And last but not least, we're gonna talk about Kubers importance in today's landscape.
  • Kubernetes is a container orchestration system. Operator is a design pattern for extending Kubernetes API. Most times operators need some, some custom resources. Let's see it in action.
  • Kubernetes API operators are custom controllers which allows us to have opinionated resources on top of Kuberneted. Now this enables us to build various things ontop of Kuobernetes. It's easy to extend its API and it's almost almost a platform for building platforms.
  • That was it. Don't forget, if you have any questions with regards to this presentation, this is my handle. Have a nice one. Thank you.


This transcript was autogenerated. To make changes, submit a PR.
Hello everybody and welcome to conference 442 Kubernetes track. If you wonder if this is yet another Kubernetes stock then the answer is maybe stick around. Then let's find out. So today's presentation is entitled I don't know kubernetes and at this point I'm too afraid to ask. This title is rather extensive and today we're gonna touch upon Kubernetes API, Kubernetes controller and operator pattern. Today's agenda is a rather simple and straightforward one. We're gonna talk a little bit about Kubernetes history, then I'm gonna share some personal blunders and aha moments. Then we're gonna have a small demo concerning operators. And last but not least, we're gonna talk about Kubernetes importance in today's landscape. A couple of words about me. My name is Alex, I'm a site reliability engineer at systematic and in my spare time I like to contribute on various platforms. You can find me under that handle Dejano Alex. Feel free to reach me if you have any questions concerning this presentation. Kubernetes has roots in two internal systems at Google. Borg, a cluster manager and the other one omega. These are some snippets from the white papers which are publicly available. I highly recommend to read them a little bit just to understand kubernetes. Now the first prototype of kubernetes was written in Java. It was actually a Borg cell running on a local machine. Later on being rewritten in go for obvious reasons. In 2014 we have the first public commit followed by one year later followed by CNCF Cloud Native Computing foundation in which Kubernetes was donated as c technology. So here we can see Google. We are pleased to contribute Kubernetes open source cluster scheduler formulation is rather interesting, right? It's not just a plain orchestrator. Now for those who don't know, CNCF Cloud Native Computing foundation is an umbrella for open source projects. This is the CNCF landscape. And in here each small square it's actually a open source technology that addresses a specific need. Here we have the main pillars, application definition and image build databases. Here we can see under scheduling and orchestration Kubernetes. And if we zoom out a little bit, if we scroll, we can see that the landscape is rather daunting. We have almost 200 technologies in CNCF. Now why it is important that Kubernetes was the c technology. Moving further, after almost ten years, our job market looks like this. Desired characteristics kubernetes technical stack kubernetes. Nice to have kubernetes. Last but not least, what your day will look like. Kubernetes. So these are snippets from actual job descriptions. And fortunately or unfortunately Kubernetes a prevalent technology. So more or less evolved under, around kubernetes, right. Being c technology in CNCF. Our entry point in today's discussion is Kubernetes Lingo, right? So let me share with you a funny story. So in the beginning, involved in various meetings, one of which someone said we need a solution that supports automatic resource bin packaging for our workloads. And I was okay, so we're not talking about Kubernetes because Kubernetes is a container orchestration system, right? But actually this is the technical definition. So coming back to the Borg white paper and, and omega, right, that, the scheduling part, the automatic resource bin packaging, this is a very important feature of kubernetes. And of course we have various terms, right? Like naked, naked pods, both which don't run under a controller. We have things like workload. So you might say my pod has been scheduled or you can say the workload has been scheduled. We have various container design patterns like init containers, sidecar containers, ambassador pattern, right? For, for containers. And also for those interested in Kubernetes administration, we have things like static pods. Now almost every presentation concerning Kubernetes starts with this high level view of Kubernetes architecture in which we have the control plane with API server, scheduler, control manager, etCD and the data plane with workers. Right? Now one might say okay, so this is the special thing concerning Kubernetes. But if we were to take a look at another technology, we can see that it also has a control plane with managers. It all has a state store, raft based state store like ETCD. And last but not least, it has a data blame with workers. So for sure we can see some similarities, right? Then this begs the question, what's the special thing about kubernetes? Even though we work with yamls on a daily basis, behind the scenes, Kubernetes has a nice HTTP based API server. So behind the scenes they're actually jsons and some, some protocol buffers, right, for internal calls. And this API server, which is the core of Kubernetes architecture implements the controller pattern, right? What do we mean by controller pattern? It's a simple watch loop that watches the desired state, meaning the one that we have in our yamls and the current state, which is the actual state of the, let's say objects that we have running in our cluster and tries to do the reconciliation, right, the famous reconciliation loop, it's actually the controller pattern which tries to adjust the current state in order to match the desired state. Now the controller mainly does three things. It observes the desired state, it analyzes, and last but not least, it corrects the drift of our current state, right? So the controller, it's, as I said, it's a, basically it, a watch loop, a generic watch loop, so to say operators. The next step, operators are a design pattern for extending Kubernetes API and creating software to run kubernetes. This is a rather steep definition. Now basically an operator is just a custom controller, right? A custom watch loop. And most times operators need some, some custom resources. Here on the right side we have some custom resource definitions. So if you do Kubectl get crds, you're gonna see custom resources that you, you are having in your cluster. Things like harbor or Jagger or keycloak configuration resources. Right. Now we're gonna try to have a demo concerning operator. Now the, the entire premise of our setup is as following, so as we said, the operator is a custom controller. We're going to have an operator running, we're going to deploy an operator in our cluster. The operator is being configured by a custom resource. So we're going to have a custom resource definition. And actually a custom resource is a way of extending Kubernetes API. And the entire purpose of this operator is to aggregate logs from various parts. So let's see it in action, starting with some, some small prerequisites. Right now in kubernetes we have the documentation at our fingertips. So if we do a Kubectl explain pod, for example, we can see exactly how the controller works, so to say, right? So we have some nested fields, the spec which is the desired state and the status, the most recently observed status of the pod, right. If we drill down a little bit and take a look at the status, we can see the all known phases like fail, pending, running, succeeded and unknown. So these are our pods, specs and status, right. The, the fields that the controller uses. Now you might have heard that kubernetes is a computer abstraction. And why is that? For example, on my local machine, if I would like to have a sorted view about my processes resource consumption, I can do a top, I will select CPU as a key for my sorting. And here we have a sorted view concerning our resource, resource consumptions for our process. But guess what? We can do the same thing in kubernetes so we do a Kubectl ted top pods. Let's take a look at the containers. Let's sort by cpu and let's take a look in all namespaces. So here we have a sorted view for our pods. Guess what? We can do the same thing for our node. So we can do Kubectl top nodes. Here I have the simple Kubernetes cluster provided by, by my setup by the Docker desktop solution. The interesting part is if we increase the verbosity a little bit, we can see that behind the scenes an API is being called, more exactly the nodes API. And we have a response body. So the JSON that we, we talked in the beginning, now we can do the same request having the endpoint, right? So we can do a row request to this endpoint and let's pipe it to JQ. And guess what we're, guess what, we're going to have a response, right? So if we do Kubectl top pods, top notch, sorry. We're going to see that we have the same response. Of course, previously we had it in nano cores and here we have millicores. Nonetheless we can talk with API. So we can see that behind the scenes. It's a nice API. Now if we take a look at APA resources, let me zoom out a little bit, we can see that we have our, let's say vanilla objects like pods, secret services deployments. Right? Here we have the APA version, the API groups, and here the names for our kinds. The kind, it's so to say object in kubernetes, more exactly, is an API resource endpoint. Now if we grab four metrics, we can see that actually this is the APA, that API endpoint that we use when we did Kubectl top pods or top nodes, right? So if we take a look at this API group, we can see that we have two objects behind matrix API group. This is the version, right? And here we have the kind meaning the objects. Now we have an idea that behind the scenes there, there's a, there's an API. We basically talk with the AP to get various information. Now the interesting part is that we can extend this API. So for example, if we would like to extend the API with so to say KCD API endpoint, we just verified we don't have it. We need something and that something is a custom resource. So if we take a look at our custom resource, there are a couple of important things in here, meaning the API group. So we said that we need or we want the KCD API group. Behind this API group we have a single object of kind log drain. Right. And last but not least, another important thing is that this object has only one spec, meaning the target spec which is a type string. So now if we were to apply this custom resource definition, we can see that now we have our new APA group of version V one alpha. And behind this API group we have this API resource kind log drain. And the nice thing is we can explain it. So if we do x explain LD, the short name logger drain, we can see that is a resource that allows the configuration of a dim operator. So it needs an operator, right? But before creating the operator, let's take a look at the specs. So if we're looking at the specs we can see that it has only one field as we we know the target field and it's a specification for managing logger aggregation to the demo operator drain. And the target is used to set the desired label view for the operator to pick up. So now this begs the question, what's an operator? We said it's a custom controller, a custom watch loop, but actually it's just a simple application, a simple binary. We can take a look, have the small script package as a weekend. Actually let's build it up. So we need to build our image, let's use the same tag and manifest. And basically we packaged our application as a docker image and now we need to deploy it. But keep in mind our application needs to talk with Kubernetes API. Therefore it needs some RBAC roles. So we need to create a service role and some cluster role bindings for our application. Of course in order to deploy our application we're gonna need to apply the deployment. So we have a deployment and if we take a look at the pods which are running in the default namespace, we can see that we have our operator which is the our application running in a pod. So if we do a Kubectl logs and follow the dim operator, we can see that our operator looks for logger drain objects. Now let's create our logger drain object. Let me split the screen here. Now of course keep in mind on the left side we have the operator, our custom loop which is running and it's looking for logger drain object. So we created our custom resource definition but we don't have any logger drain objects. So if we do a Kubectl get logger drain, there's nothing. Now let's create an object. So let's create our local drain object. It's of kind. Log drain. Let's call it demo LD. And here let's set the target as being KCD. And now as soon as we create our logger drain object, so we're going to kubectl apply it, we can say that our pod has already seen the logger drain object, right? So logger drain object named demo LD with specs target KCD. So that's nice. Now we have a logger drain object which tells to our operator pick up the logs for all ports that have the label KCD. But we don't have any pods label kcd with the label KCD. Therefore we need to create. I'm going to use a simple Nginx image and I'll create a naked pod. And as soon as my nginx pod named pod one will be healthy and ready, we can see that our operator picked up its log so it found pod one in default namespaces. And here we can see the logs of the nginx based pod. We can make another naked pod. Remember, the scope of the operator was to drain the logs for all pods with the label target KCD. So here it's a simple flask based application and the story remains the same. As soon as the board is healthy, we can see that the logs have been picked up for pod two as well. And here we can see the pods right now. Coming back to our presentation, what we've seen, we've seen that a resource in Kubernetes is actually an endpoint, right? It's an endpoint. Kubernetes API operators are custom controllers which allows us to have opinionated resources on top of Kubernetes, meaning that we have a way of extended Kubernetes API. Now this enables us to build various things on top of Kubernetes, right? And here we've seen the entire adoption of platform engineering, right? It's rather easy to extend Kubernetes API, but this doesn't mean that you really need to do it. I mean you can have something that it works, but it lacks a proper developer experience or a proper use case, right? So keep in mind, if you want to extend Kubernetes API is rather easy using custom resources and custom controllers, think twice before doing it, right? In many ways, everything is an abstraction, right? So we're working with abstraction on daily basis, right? So when you create a custom resource in an operator, you create an abstraction that can be used in Kubernetes. As we've seen, Kubernetes is a platform for building platforms, right? It's easy to extend its API and it has almost of a plug and play approach with regards to the API. And last but not least, a very important thing. Kubernetes is only as good as the infrastructure it runs on top of, right? So keep in mind having multiple small nodes versus a few larger nodes. This is a ongoing debate and Kubernetes will not do any magic if the infrastructure under provision. That was it. Thank you very much. Don't forget, if you have any questions with regards to this presentation, this is my handle. Have a nice one. Goodbye. Thank you.

Alex Dejanu

SRE @ Systematic

Alex Dejanu's LinkedIn account Alex Dejanu's twitter account

Awesome tech events for

Priority access to all content

Video hallway track

Community chat

Exclusive promotions and giveaways