It took me a long time to wrap my head around the idea of Kubernetes. Even if I’ve done some projects using k8s before, there’s still this voice at the back of my head saying: “Ok, I really don’t know how this whole thing works.”
In this blogpost, I’ll talk about my journey in understanding Kubernetes—where learning Pods and Nodes proved to be a false start—and why I think that Deployments and Services is a good place to begin. Lastly, I’d argue that k8s may be better taught first with the concept of Deployments and Services!1
This post is divided into two parts:
- Why starting with Pods and Nodes didn’t help me
- Why starting with Deployments and Services clicked right away
Before we begin, a little something about myself: before learning Kubernetes, I’m already familiar with container tools such as Docker and docker-compose. I’ve built simple web applications before and deployed them to platforms like Heroku or App Engine so I kinda get the idea of “putting something in the internet.” Perhaps this has affected my learning trajectory for k8s, and might be different for you or your team.
Lastly, this is not a tutorial about Deployments & Services nor Kubernetes as a whole. Here, I’ll try to be reflective but informative. I may be explaining some terminologies along the way, but will pull-in Kubernetes concepts here and there.
Pods are simple at first, but overloads you after a while
Most tutorials I’ve seen in the Internet starts with a Pod, and for good reason: because you can think of Pods as a basic unit in Kubernetes. In many cases, a Pod is often mapped to a container—and hey, Kubernetes is a container orchestration tool!
Ok, that’s cool! I can now map what I know to this new Kubernetes concept! So where did the confusion began? It began when I saw notes like these:
“Pods (can) run multiple containers that need to work together.” (“Understanding Pods”, Kubernetes Documentation)
It’s not bad, but it sets off a conceptual alarm that Pods may not be as straightforward as what they appear to be. This proved to be true because in practice, you don’t directly work with Pods, you delegate a Controller to manage them for you.
“Kubernetes uses a higher-level abstraction, called a Controller, that handles the work of managing the relatively disposable Pod instances. Thus, while it is possible to use Pod directly, it’s far more common in Kubernetes to manage your pods using a Controller.” (“Working with Pods”, Kubernetes Documentation)
(With the gift of hindsight, this statement hinted the idea of Deployments.)
I think the main reason why Pods are a poor starting point in learning Kubernetes is because it does not provide an “actionable” mapping of what I already know and what I’m trying to learn. Even if I now know that a Pod is a basic unit and it maps to a container, it is not actionable enough that I can apply this knowledge when using Kubernetes. On another note, starting from Pods will inadvertently add another hoop to jump into— I need to now know what Controllers are, and there’s three of them for my novice mind to learn.
Nodes pull you into the architecture rabbit-hole
Just like Pods, Nodes are unassuming: they’re just worker machines where processes are run. The documentation says that they may be “a VM or a physical machine depending on the cluster.” As someone who used to SSH into VMs and run Python programs, Nodes may be a relatable starting point.
However, as I scrolled through the documentation, I saw things about external
IPs, memory, PID pressure, and whatnot. Upon seeing these words, I was
intimidated. Wait! I thought one benefit of “Serverless and The Cloud” is
that I don’t need to think about what’s happening in my machine? The concept
of Nodes got a bit too low-level right away that I was sucked into the
kubelet rabbit-hole and enrolled myself into a “Learn Kubernetes the Hard
Way” course in
LinuxAcademy (which I wasn’t able to finish by the
The problem with Nodes is that they are too low-level of an abstraction for me to get started. Even if it’s a natural progression from virtual machines conceptually, it didn’t lead me to do what I need to do, that is, to just deploy an application.
Some things are good conceptually but bad practically
Understanding Pods and Nodes are good conceptually, since they are a natural progression of what I know so far: Docker containers can be mapped into Pods, and the machines where these containers are run can be mapped into Nodes. However in practice, this bottom-up approach didn’t help because there’s more to Kubernetes than Pods and Nodes. If I want to deploy an application, there’s a lot of things I still need to think about.
In the next section, I’ll discuss why the concept of Deployments and Services provide the “right amount of abstraction” in order for me to (1) understand what’s happening and (2) be productive with it right away. On this level, we can even consider Deployments as a functional unit of a Kubernetes cluster.
By first learning about deployments and services, I immediately had a handle on the simplest and oft-used Kubernetes use-case: “I want to have X in my cluster.”
The simplest and oft-used Kubernetes use-case: “I want to have X in my cluster”
# "I want to have Redis in my cluster" apiVersion: apps/v1 kind: Deployment spec: # ... template: spec: containers: # Make me a container... - name: redis # ...named `redis` based on... image: redis # ..this image
In contrast to the Pods analogy, Deployments provide a one-to-one and actionable conceptual mapping of “what you want to do” and “how you can do it.” Every time I want to have X in my cluster, I can just think of any Docker image at my disposal (e.g. a machine learning service, a node backend, or a Postgres database).
Next, we can incorporate the idea of Services when we expand the statement above into: “I want to have X in my cluster that is accessible via Y.” Kubernetes is a clever way of orchestrating processes and networks together where Deployments fill-in the former and Services the latter.
“I want to have X in my cluster that is accessible via Y.”
So if we want to say: “I want to have Redis in my cluster that is accessible only within the cluster”, then we just map this sentence into a configuration file consisting of a Deployment and a Service3:
# "I want to have <X> Redis in my cluster that is accessible # <Y> only within the cluster" apiVersion: apps/v1 kind: Deployment spec: # ... selector: matchLabels: app: redis template: spec: containers: # Make me a container... - name: redis # ...named `redis` based on... image: redis # ..this image --- apiVersion: v1 kind: Service spec: type: ClusterIP ports: - port: 6379 targetPort: 6379 selector: app: redis
By thinking in terms of Deployments and Services, it is easier to transform a declaration (“I want…“) into a configuration file. In my experience, this conceptual framework helped me grok the concept of Kubernetes.
Why Deployments and Services clicked
There are three reasons why the concepts of deployments and services easily clicked for me:
- It is a direct mapping between “what I want to do” and “how to do it”. As we have seen from the previous examples, it is straightforward and there are no conceptual hoops to jump on to. As long as I can declare “I want X in my cluster,” I can conveniently transform that into a k8s configuration file.
- It is an easy next-step after learning about Containers. Deployments only care about the container image you have—plus some other settings that you may or may not use. In order to be productive in Kubernetes, I just need to know what those settings are (replicas, labels, etc.). It’s a small delta between Kubernetes and Containers, and it made me productive in k8s immediately.
- It is easy to onboard people to your infrastructure. Once I start with Deployments and Services, I found it easier to onboard people in my cluster setup. I just show them the images I’ve built and the deployment configuration file. Then once I’ve established their connection, they can now work with my cluster immediately.
In this blogpost, I talked about my journey in learning Kubernetes, and how learning about Deployments and Services made k8s concepts suddenly click. First, I illustrated how Pods and Nodes proved to be false starts: Pods have “hidden concepts” that overloads you after a while, and Nodes can bring you into an architecture rabbit-hole. Then, I talked about Deployments and Services and why we should start with them. It clicked because:
- It provides a direct mapping of what I want to do and how to do it
- It is an easy next-step after learning about Containers, and
- It is easy to onboard people into my cluster setup
Learning Kubernetes was really challenging and fun. It seems that after I grokked these concepts, I unlocked a whole new world of possibilities! This made me more excited to work on distributed systems and infrastructure. Wish me luck!
- 03-04-2020: Update and flesh-out the conclusion and why it clicked