Chapter 18: Kubernetes Workloads

18.1 Why Kubernetes — Compose does not survive a dead host

By Ravindra Bagale · Last updated: 25 October 2026

Why: the problem

Think of Swiggy on a Friday night. The order service runs in Docker Compose on one EC2 server. At 8:30 p.m. that server's disk fails. Compose does not move anything. Every container was on that one machine, so every container is gone. Orders stop until someone wakes up, makes a new server and runs docker compose up by hand. That can take 40 minutes. On a busy night that costs real money. Compose is great on one machine. It does not know that other machines exist. Kubernetes fixes this. You give it several machines, called nodes. You tell it what you want, for example three copies of the order service. If a node dies, Kubernetes notices within about a minute and starts the lost copies on the nodes that are still healthy. Nobody has to wake up. In this lesson you will see this happen for real. You will run the same small app in Compose and in a three-node kind cluster, then kill a node and watch what each tool does.

What we will build

You will run a small nginx web app two ways. First with Docker Compose on your laptop. Then as a Deployment with three replicas on a three-node kind cluster. Then you will stop one node container and watch Kubernetes move the Pods to the other nodes, while Compose has no answer for a dead host.

What: every term in plain English

• Node — One machine (VM or physical) that runs your containers. In kind, each node is a Docker container.
• Cluster — A group of nodes managed together by Kubernetes.
• Control plane — The brain of the cluster. It stores what you asked for and makes it happen.
• Worker node — A node that runs your app Pods.
• Desired state — What you asked for, like "3 copies of nginx".
• Actual state — What is really running right now.
• Reconcile — The control loop that keeps fixing actual state until it matches desired state.
• kind — Kubernetes IN Docker. A free tool that makes a real cluster out of Docker containers for practice.
• kubectl — The command line tool you use to talk to the cluster.

How: step by step

1. Write a three-node kind config

   cat kind-3.yaml
   

   What it does: A config file that asks kind for one control plane and two workers.

   Output

   kind: Cluster
   apiVersion: kind.x-k8s.io/v1alpha4
   nodes:
   - role: control-plane
   - role: worker
   - role: worker
   

2. Create the cluster

   kind create cluster --name lab --config kind-3.yaml
   

   What it does: Starts three node containers and joins them into one cluster.

   Output

   Creating cluster "lab" ...
    ✓ Preparing nodes 📦 📦 📦
    ✓ Joining worker nodes 🚜
   Set kubectl context to "kind-lab"
   

3. Check the nodes

   kubectl get nodes -o wide
   

   What it does: Lists every node, its role and its internal IP.

   Output

   NAME                STATUS   ROLES           AGE   VERSION   INTERNAL-IP
   lab-control-plane   Ready    control-plane   60s   v1.30.0   172.18.0.4
   lab-worker          Ready    none            40s   v1.30.0   172.18.0.3
   lab-worker2         Ready    none            40s   v1.30.0   172.18.0.2
   

4. Run three copies of nginx

   kubectl create deployment web --image=nginx:1.27 --replicas=3
   

   What it does: Asks Kubernetes for three nginx Pods. It picks the nodes itself.

   Output

   deployment.apps/web created
   

5. See where the Pods landed

   kubectl get pods -o wide
   

   What it does: The NODE column shows the Pods spread across the workers.

   Output

   NAME                   READY   STATUS    NODE
   web-6d4cf56db6-2xkqp   1/1     Running   lab-worker
   web-6d4cf56db6-8m9tz   1/1     Running   lab-worker2
   web-6d4cf56db6-q7rvn   1/1     Running   lab-worker
   

6. Kill a node

   docker stop lab-worker
   

   What it does: Stops the node container. This is the same as a server losing power.

   Output

   lab-worker
   

7. Watch Kubernetes heal

   kubectl get pods -o wide -w
   

   What it does: After about a minute the node is NotReady and after the eviction timeout new Pods start on lab-worker2.

   Output

   web-6d4cf56db6-2xkqp   1/1   Terminating   lab-worker
   web-6d4cf56db6-zp4lc   0/1   Pending       none
   web-6d4cf56db6-zp4lc   1/1   Running       lab-worker2
   

Example 1: The app in Docker Compose, on one host

Stack. Docker Compose v2, nginx 1.27

Goal. The app in Docker Compose, on one host.

compose.yaml

services:
  web:
    image: nginx:1.27
    ports:
      - "8080:80"
    deploy:
      replicas: 1
    restart: always

terminal

docker compose up -d
curl -s -o /dev/null -w "%{http_code}\n" localhost:8080
docker compose ps

What each line does

• image: nginx:1.27 — A pinned nginx version, never latest.
• ports 8080:80 — Laptop port 8080 goes to container port 80.
• restart: always — Docker restarts the container if it crashes, but only on this same host.
• docker compose ps — Shows the container and its state.

Green output

[+] Running 2/2
 ✔ Network demo_default  Created
 ✔ Container demo-web-1  Started
200
NAME         IMAGE        STATUS
demo-web-1   nginx:1.27   Up 5 seconds

Expected result. The app works. But restart: always only helps when the process crashes. If the whole host dies, nothing restarts it anywhere else.

Example 2: The same app as a Kubernetes Deployment with 3 replicas

Stack. kind v0.23, Kubernetes 1.30

Goal. The same app as a Kubernetes Deployment with 3 replicas.

web.yaml

apiVersion: apps/v1
kind: Deployment
metadata:
  name: web
spec:
  replicas: 3
  selector:
    matchLabels:
      app: web
  template:
    metadata:
      labels:
        app: web
    spec:
      containers:
      - name: nginx
        image: nginx:1.27
        ports:
        - containerPort: 80

terminal

kubectl apply -f web.yaml
kubectl rollout status deployment/web
kubectl get pods -l app=web -o wide

What each line does

• kind: Deployment — The object that keeps N copies of a Pod running.
• replicas: 3 — The desired state: three Pods.
• selector.matchLabels — How the Deployment finds its own Pods.
• template — The Pod recipe. Every copy is made from this.
• labels app: web — The tag that must match the selector.
• kubectl rollout status — Waits until all three are ready.

Green output

deployment.apps/web created
Waiting for deployment "web" rollout to finish: 1 of 3 updated replicas are available...
deployment "web" successfully rolled out
NAME                   READY   STATUS    NODE
web-7c9d8b6f5d-4lq2m   1/1     Running   lab-worker
web-7c9d8b6f5d-b8wxs   1/1     Running   lab-worker2
web-7c9d8b6f5d-k2r9d   1/1     Running   lab-worker

Expected result. Three copies run on two different nodes. You wrote what you want, not where to run it.

Example 3: The dead host test: Compose versus Kubernetes

Stack. kind 3 nodes, Docker 26

Goal. The dead host test: Compose versus Kubernetes.

terminal

# Compose: stop the only host (here, the Docker engine)
sudo systemctl stop docker
curl -s -m 3 localhost:8080 || echo "site down"
sudo systemctl start docker

# Kubernetes: stop one worker node
docker stop lab-worker
kubectl get nodes
sleep 330
kubectl get pods -l app=web -o wide

What each line does

• systemctl stop docker — Pretends the Compose host died. All containers stop.
• docker stop lab-worker — Pretends one Kubernetes node died.
• kubectl get nodes — Shows lab-worker as NotReady after about 40 seconds.
• sleep 330 — The default eviction wait is 5 minutes. After it, Pods on the dead node are replaced.
• get pods -o wide — All 3 Pods now run on lab-worker2.

Green output

site down
NAME                STATUS     ROLES           AGE
lab-control-plane   Ready      control-plane   12m
lab-worker          NotReady   none            12m
lab-worker2         Ready      none            12m
NAME                   READY   STATUS    NODE
web-7c9d8b6f5d-b8wxs   1/1     Running   lab-worker2
web-7c9d8b6f5d-n5t7h   1/1     Running   lab-worker2
web-7c9d8b6f5d-x9pz2   1/1     Running   lab-worker2

Expected result. Compose: the site is down until a human acts. Kubernetes: the desired state (3) is restored on the healthy node without anyone typing a command.

Verify: did it work?

1. Count ready replicas

   kubectl get deployment web
   

   What it does: READY must say 3/3 again after the healing.

   Output

   NAME   READY   UP-TO-DATE   AVAILABLE   AGE
   web    3/3     3            3            14m
   

2. See the events

   kubectl get events --sort-by=.lastTimestamp | tail -5
   

   What it does: Shows the node going NotReady and new Pods being scheduled.

   Output

   Warning  NodeNotReady  node/lab-worker  Node lab-worker status is now: NodeNotReady
   Normal   Scheduled     pod/web-...-n5t7h  Successfully assigned practice/web-...-n5t7h to lab-worker2
   

When it is red 1

ERROR: failed to create cluster: node(s) already exist for a cluster with the name "lab"

Cause. A cluster called lab is already there from an old lesson.

Fix, step by step

1. Run kind get clusters.
2. Delete it with kind delete cluster --name lab, or use another name.
3. Create again.

When it is red 2

The connection to the server localhost:6443 was refused - did you specify the right host or port?

Cause. The cluster is not running, or kubectl points to an old context.

Fix, step by step

1. Run kubectl config current-context.
2. Run docker ps and check the lab-control-plane container is Up.
3. Run kubectl config use-context kind-lab.

When it is red 3

Pods stay Running on lab-worker even after docker stop

Cause. You looked too early. Kubernetes waits for the node to be NotReady and then for the eviction time (about 5 minutes).

Fix, step by step

1. Run kubectl get nodes until lab-worker is NotReady.
2. Wait 5 more minutes.
3. Check again with kubectl get pods -o wide.

Cleanup

1. Bring the node back

   docker start lab-worker
   

   What it does: The node rejoins as Ready. Kubernetes does not move Pods back by itself.

   Output

   lab-worker
   

2. Delete the app

   kubectl delete deployment web
   

   What it does: Removes the Deployment and its Pods.

   Output

   deployment.apps "web" deleted
   

3. Delete the lab cluster

   kind delete cluster --name lab
   

   What it does: Removes the whole cluster container. On minikube use: minikube delete.

   Output

   Deleting cluster "lab" ...
   Deleted nodes: ["lab-control-plane"]
   

4. Stop Compose

   docker compose down
   

   What it does: Removes the Compose container and network.

   Output

   [+] Running 2/2
    ✔ Container demo-web-1  Removed
   

Practice tasks

Homework: try at home

Recap

• Compose runs on one host and cannot survive that host dying.
• Kubernetes runs on many nodes and keeps desired state.
• You write what you want; the scheduler picks where.
• kind gives you a real multi-node cluster on a laptop.
• Healing after a node loss takes about 5 minutes by default.

FAQ

Do I need Kubernetes for a small project?

No. One server with Compose is fine for small apps. Use Kubernetes when you need many servers, self-healing and zero-downtime updates.

Is kind good for production?

No. kind is for learning and CI tests. Production uses EKS, GKE, AKS or k3s on real servers.

kind or minikube?

Both are fine. kind is faster and supports many nodes easily. minikube has handy addons like ingress and dashboard.

Does Kubernetes replace Docker?

No. Kubernetes runs containers. The images you build with Docker still run on it.

Why did the Pods not move back when the node came back?

Kubernetes does not rebalance by itself. New Pods may go to that node later.

Interview questions

1. Why would a company move from Compose to Kubernetes?

   To run across many servers, survive a server dying, scale up and down, and do rolling updates without downtime.

2. What is desired state?

   The state you write in YAML, like 3 replicas. Controllers keep changing the cluster until it matches.

3. What happens when a node dies?

   The node turns NotReady, and after the eviction timeout the Pods there are replaced on healthy nodes by their controller.

4. Can a bare container restart policy replace this?

   No. restart: always only works on the same host. It cannot move work to another machine.

5. How do you practise Kubernetes without cloud cost?

   With kind or minikube on a laptop.

Samajla ka?

Ravi Sir's Cards for this lesson

Revision cardInterview card

मराठी cards