K8S Deployment

Skill — Kubernetes Deployment

When this skill activates

Any task involving Kubernetes deployments: writing or modifying manifests, Helm charts, configuring autoscaling, probes, network policies, ingress, resource management, or deployment strategies for containerized workloads.

Mandatory actions when this skill is active

Before writing any code

1. Identify the deployment requirements:
   • Target environment (dev/staging/production)
   • Availability requirements (SLA target, max acceptable downtime)
   • Scale expectations (baseline replicas, peak load multiplier)
   • Network exposure (internal only, public ingress, specific CIDR allowlists)
2. Check existing cluster state:

   kubectl get nodes -o wide          # Cluster capacity
   kubectl top nodes                   # Current resource usage
   kubectl get namespaces              # Available namespaces
   kubectl get resourcequotas -A       # Existing quotas
   

3. Determine if Helm or raw manifests are appropriate:
   • Helm: Multiple environments, parameterized configs, community chart availability.
   • Raw manifests + Kustomize: Simpler apps, GitOps with ArgoCD/Flux, overlay-based config.

During implementation

Deployment Strategies

• RollingUpdate (default, recommended for most services):

  apiVersion: apps/v1
  kind: Deployment
  metadata:
    name: my-service
  spec:
    replicas: 3
    strategy:
      type: RollingUpdate
      rollingUpdate:
        maxSurge: 1          # Max extra pods during update
        maxUnavailable: 0    # Zero downtime: never kill before new is ready
    selector:
      matchLabels:
        app: my-service
    template:
      metadata:
        labels:
          app: my-service
      spec:
        containers:
        - name: my-service
          image: registry.example.com/my-service:v1.2.3
          # ... rest of spec
  

• Recreate (only for stateful apps that cannot run two versions simultaneously):

  strategy:
    type: Recreate
  

• Blue/Green (via service selector swap or Argo Rollouts): Deploy new version as separate deployment, switch service selector when healthy.
• Canary (via Argo Rollouts or Istio traffic splitting): Route percentage of traffic to new version, promote or rollback based on metrics.

Helm Chart Structure

my-chart/
  Chart.yaml          # Chart metadata (name, version, appVersion)
  values.yaml         # Default configuration values
  values-staging.yaml # Environment-specific overrides
  values-prod.yaml
  templates/
    _helpers.tpl      # Template helpers and labels
    deployment.yaml
    service.yaml
    hpa.yaml
    ingress.yaml
    configmap.yaml
    secret.yaml       # Reference to external secrets, not raw values
    pdb.yaml
    networkpolicy.yaml
    serviceaccount.yaml
  tests/
    test-connection.yaml

• Chart.yaml must include:

  apiVersion: v2
  name: my-service
  version: 1.0.0       # Chart version (bump on chart changes)
  appVersion: "1.2.3"  # Application version (matches container tag)
  

• values.yaml conventions:

  replicaCount: 3
  image:
    repository: registry.example.com/my-service
    tag: ""            # Overridden per environment
    pullPolicy: IfNotPresent
  resources:
    requests:
      cpu: 100m
      memory: 128Mi
    limits:
      cpu: 500m
      memory: 512Mi
  

HPA (Horizontal Pod Autoscaler)

apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: my-service
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: my-service
  minReplicas: 3         # Never go below 3 for production
  maxReplicas: 20        # Cap to prevent runaway scaling
  metrics:
  - type: Resource
    resource:
      name: cpu
      target:
        type: Utilization
        averageUtilization: 70    # Scale up at 70% CPU
  - type: Resource
    resource:
      name: memory
      target:
        type: Utilization
        averageUtilization: 80
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 300   # Wait 5min before scaling down
      policies:
      - type: Percent
        value: 25
        periodSeconds: 60               # Max 25% reduction per minute
    scaleUp:
      stabilizationWindowSeconds: 30
      policies:
      - type: Percent
        value: 100
        periodSeconds: 30               # Can double pods in 30s

• Custom metrics (requests per second, queue depth) via Prometheus adapter:

  - type: Pods
    pods:
      metric:
        name: http_requests_per_second
      target:
        type: AverageValue
        averageValue: "1000"
  

Pod Disruption Budget (PDB)

apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: my-service-pdb
spec:
  minAvailable: 2        # Always keep at least 2 pods running
  # OR: maxUnavailable: 1  # At most 1 pod can be down
  selector:
    matchLabels:
      app: my-service

• Always create a PDB for production workloads. Without one, cluster upgrades and node drains can take down all pods simultaneously.
• Rule of thumb: minAvailable = replicas - 1 or use maxUnavailable: 1.

Probes (Health Checks)

containers:
- name: my-service
  livenessProbe:           # Is the process alive? Restart if failing.
    httpGet:
      path: /healthz
      port: 8080
    initialDelaySeconds: 15
    periodSeconds: 10
    failureThreshold: 3    # 3 failures = restart
    timeoutSeconds: 3
  readinessProbe:          # Is it ready for traffic? Remove from LB if failing.
    httpGet:
      path: /readyz
      port: 8080
    initialDelaySeconds: 5
    periodSeconds: 5
    failureThreshold: 2
    timeoutSeconds: 3
  startupProbe:            # For slow starters. Disables liveness/readiness until passing.
    httpGet:
      path: /healthz
      port: 8080
    initialDelaySeconds: 0
    periodSeconds: 5
    failureThreshold: 30   # 30 * 5s = 150s max startup time
    timeoutSeconds: 3

• liveness: "Is the process stuck?" Triggers container restart. Keep simple (not DB-dependent).
• readiness: "Can it serve requests?" Controls load balancer membership. Can check dependencies.
• startup: Use for apps that take > 30s to initialize (JVM warmup, large model loading).

Resource Requests and Limits

resources:
  requests:              # Scheduling guarantee (must be available)
    cpu: 100m           # 0.1 CPU cores
    memory: 256Mi       # 256 MiB RAM
  limits:               # Hard ceiling (OOMKilled if exceeded for memory)
    cpu: 1000m          # 1 CPU core (throttled, not killed)
    memory: 512Mi       # OOMKilled if exceeded

• Requests: Set to observed p50 usage. Cluster scheduler uses this for placement.
• Limits: Set to observed p99 + 20% headroom. Too tight = OOMKills. Too loose = noisy neighbors.
• CPU limits debate: Some teams remove CPU limits (use only requests) to avoid throttling. This is acceptable if the cluster has sufficient headroom and resource quotas protect namespaces.
• Always set memory limits (OOM without limits can crash the node).

Network Policies

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: default-deny-all
  namespace: my-namespace
spec:
  podSelector: {}          # Applies to all pods in namespace
  policyTypes:
  - Ingress
  - Egress
---
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
  name: allow-my-service-ingress
  namespace: my-namespace
spec:
  podSelector:
    matchLabels:
      app: my-service
  policyTypes:
  - Ingress
  ingress:
  - from:
    - namespaceSelector:
        matchLabels:
          name: ingress-nginx
    ports:
    - protocol: TCP
      port: 8080

• Default deny first, then allow specific traffic paths.
• Minimum for production: deny all ingress/egress, then whitelist:
  1. Ingress controller to service pods.
  2. Service pods to database pods.
  3. Egress to external APIs (specific IPs/CIDRs if possible).

Ingress Configuration

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: my-service-ingress
  annotations:
    cert-manager.io/cluster-issuer: letsencrypt-prod
    nginx.ingress.kubernetes.io/rate-limit: "100"
    nginx.ingress.kubernetes.io/rate-limit-window: "1m"
spec:
  ingressClassName: nginx
  tls:
  - hosts:
    - api.example.com
    secretName: api-tls-cert
  rules:
  - host: api.example.com
    http:
      paths:
      - path: /
        pathType: Prefix
        backend:
          service:
            name: my-service
            port:
              number: 80

• Always configure TLS termination (use cert-manager for automatic certificate management).
• Add rate limiting annotations to prevent abuse.
• Use path-based routing to split traffic to different services under one domain.

After implementation

1. Validate manifests before applying:

   helm template my-chart ./chart -f values-prod.yaml | kubectl apply --dry-run=client -f -
   kubectl diff -f manifest.yaml    # Show what would change
   

2. Verify rollout health:

   kubectl rollout status deployment/my-service --timeout=300s
   kubectl get pods -l app=my-service -o wide
   kubectl top pods -l app=my-service
   

3. Test probes manually:

   kubectl exec -it <pod> -- curl -s localhost:8080/healthz
   kubectl exec -it <pod> -- curl -s localhost:8080/readyz
   

4. Verify network policies:

   # From a test pod, confirm blocked traffic is actually blocked
   kubectl run test --rm -it --image=busybox -- wget -qO- --timeout=3 http://my-service:8080
   

5. Test PDB during a drain:

   kubectl drain <node> --ignore-daemonsets --dry-run=client
   

Common mistakes to flag

• No PDB on production deployments (cluster upgrades will cause downtime).
• Liveness probe checks database connectivity (cascading restarts on DB issues).
• No resource limits (one pod can starve the entire node).
• latest tag in production (non-reproducible deployments).
• Secrets in ConfigMaps or values.yaml (use sealed-secrets, external-secrets, or vault).
• No network policies (all pods can communicate with all other pods by default).
• HPA and VPA both active on the same resource (they conflict).

Self-check before task completion

Before marking a task done when this skill was active:

• All manifests pass kubectl apply --dry-run=client.
• Deployment has both readiness and liveness probes configured.
• Resource requests and memory limits are set on all containers.
• PDB exists for production deployments (minAvailable or maxUnavailable).
• Network policies enforce least-privilege communication.
• No secrets stored in plain text in manifests or values files.
• Image tags are pinned to specific versions (not latest).
• HPA configured with appropriate min/max and scale-down stabilization.
• Ingress has TLS termination configured.
• Rollout tested with kubectl rollout status.