statefulset.yaml

#
#  Author: Hari Sekhon
#  Date: [% DATE  # 2019-11-28 18:51:59 +0000 (Thu, 28 Nov 2019) %]
#
#  vim:ts=2:sts=2:sw=2:et
#  lint: k8s
#
#  https://github.com/HariSekhon/Kubernetes-configs
#
#  License: see accompanying Hari Sekhon LICENSE file
#
#  If you're using my code you're welcome to connect with me on LinkedIn
#  and optionally send me feedback to help improve or steer this or other code I publish
#
#  https://www.linkedin.com/in/HariSekhon
#

# ============================================================================ #
#                             S t a t e f u l S e t
# ============================================================================ #

# https://kubernetes.io/docs/concepts/workloads/controllers/statefulset/
#
# https://kubernetes.io/docs/reference/kubernetes-api/workload-resources/stateful-set-v1/
#
# see pod anti-affinity here:
#
# https://kubernetes.io/docs/tutorials/stateful-application/zookeeper/

---
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: NAME
  namespace: NAMESPACE
  annotations:
    # see also k8s_pod_disruption_budget.yaml
    # XXX: set to true if this can be evicted, otherwise this can prevent cluster autoscaler from scaling down the nodepool
    cluster-autoscaler.kubernetes.io/safe-to-evict: "false"
    # adds ArgoCD link symbol in the UI next to object
    #link.argocd.argoproj.io/external-link: http://www.domain.com/path
spec:
  replicas: 3  # XXX: don't define replicas if using HPA, otherwise gets overridden and reset to this number every deploy which could cause load or data sharding issues
  serviceName: NAME
  selector:
    matchLabels:
      app: APP
  template:
    metadata:
      labels:
        app: APP
    spec:
      # for business critical workloads to get priority access to the stable node pool instead of preemptible node pool, will evict lower priority pods to preemptible node pool (default 0) if necessary. Requires: priorityclass.yaml
      #
      # priorityClassName: high-priority
      #
      #nodeSelector:
        #myLabel: myFastNodes
        #disktype: ssd
        #kubernetes.io/arch: amd64
        #kubernetes.io/os: linux
        #node.kubernetes.io/instance-type: m3.medium    # 1.17+
        ## for persistent disks limited to 2 zones combined with GKE regional clusters
        #topology.kubernetes.io/zone: europe-west2-a  # 1.17+
        #topology.kubernetes.io/region: europe-west2    # 1.17+
        #
        # GKE specific - hard requirement - prefer affinity below is safer soft requirement to still allow scheduling if node pool becomes unavailable (eg. recreated by Terraform)
        #
        # cloud.google.com/gke-nodepool: POOL_NAME
        #
        # GKE Autopilot
        #
        #   https://cloud.google.com/kubernetes-engine/docs/concepts/autopilot-compute-classes#when-to-use
        #
        # cloud.google.com/compute-class: Balanced
        #
      #
      # https://kubernetes.io/docs/concepts/configuration/assign-pod-node/#affinity-and-anti-affinity
      # XXX: not recommended in clusters greater than several hundred nodes due to significant CPU overhead slowing down scheduling
      affinity:
        nodeAffinity:
          # Node preemption:
          # - shuts down pods ungracefully
          # - ignores PodDisruptionBudget
          # - invalidate Stateful guarantees and can lead to data loss !!
          # https://cloud.google.com/kubernetes-engine/docs/how-to/preemptible-vms#kubernetes_constraint_violations
          # https://cloud.google.com/kubernetes-engine/docs/how-to/preemptible-vms#best_practices
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
              - matchExpressions:
                  # GCP: https://cloud.google.com/kubernetes-engine/docs/how-to/preemptible-vms#best_practices
                  #
                  # gke-preemptible not valid on GKE AutoPilot clusters - remove in that case and leave only gke-spot
                  - key: cloud.google.com/gke-preemptible
                    operator: DoesNotExist
                  - key: cloud.google.com/gke-spot
                    operator: DoesNotExist
                  #
                  # AWS: https://docs.aws.amazon.com/eks/latest/userguide/managed-node-groups.html
                  - key: eks.amazonaws.com/capacityType
                    operator: NotIn
                    values:
                      - SPOT
                  #
                  # Azure: https://docs.microsoft.com/en-us/azure/aks/spot-node-pool
                  - key: kubernetes.azure.com/scalesetpriority
                    operator: NotIn
                    values:
                      - spot
        #podAffinity:
        podAntiAffinity:
          # XXX: spread app across AZs
          # XXX: use preferred if num_replicas > num_AZs
#          preferredDuringSchedulingIgnoredDuringExecution:
#            - weight: 100
#              podAffinityTerm:
#                topologyKey: topology.kubernetes.io/zone
#                labelSelector:
#                  matchExpressions:
#                    - key: app
#                      operator: In
#                      values:
#                        - APP
          requiredDuringSchedulingIgnoredDuringExecution:
            # XXX: spread app across AZs
            - topologyKey: topology.kubernetes.io/zone
              labelSelector:
                matchExpressions:
                  - key: app
                    operator: In
                    values:
                      - APP
            # XXX: spread app across hosts (default prefers spread, whereas this requires it)
            - topologyKey: kubernetes.io/hostname
              labelSelector:
                matchExpressions:
                  - key: app
                    operator: In
                    values:
                      - APP
      #serviceAccountName: NAME
      securityContext:
        # see pod-security-policy.yaml for wider enforcement
        runAsNonRoot: true
        #runAsUser: 0   # not recommended
        fsGroup: 1000   # group to own volumeMounts so they can be written to by the user-level process
        # XXX: don't chgrp all the many small files unless the root mismatches, prevents long outages on pod creation hanging with multiple of these error messages:
        #
        #   Unable to attach or mount volumes: unmounted volumes=[jenkins-home], unattached volumes=[kube-api-access-6q2kb jenkins-home]: timed out waiting for the condition
        #
        #   https://docs.cloudbees.com/docs/cloudbees-ci-kb/latest/cloudbees-ci-on-modern-cloud-platforms/timeout-when-attaching-volumes-in-kubernetes
        #
        fsGroupChangePolicy: OnRootMismatch
      containers:
        - name: NAME
          image: postgres:9.6
          ports:
            - name: postgres
              containerPort: 5432
              protocol: TCP
          securityContext:
            runAsNonRoot: true
            privileged: false
            allowPrivilegeEscalation: false
            readOnlyRootFilesystem: true
            capabilities:
              drop:
                - ALL
          env:
            - name: HEAP_SIZE
              value: 1G
            - name: CASSANDRA_SEEDS
              value: cassandra-0.cassandra.svc.cluster.local,cassandra-1.cassandra.svc.cluster.local,cassandra-2.cassandra.svc.cluster.local
            - name: POD_IP
              valueFrom:
                fieldRef:
                  fieldPath: status.podIP  # K8S API path
            - name: CLOUD_SQL_INSTANCE
              valueFrom:
                configMapKeyRef:
                  key: cloud-sql-instance
                  name: my-configmap
            - name: MYSECRET
              valueFrom:
                secretKeyRef:
                  name: my-secret
                  key: my-secret
          resources:
            requests:
              cpu: 100m
              memory: 128Mi
            limits:
              cpu: "1"
              memory: 2Gi
          readinessProbe:
            # http, tcp or exec
            #httpGet:
            #  path: /healthz
            #  port: 8080
            #  scheme: HTTPS
            # only uncomment if you can't use http check, must comment/delete http check otherwise will get this error:
            # invalid: spec.containers[0].livenessProbe.tcpSocket: Forbidden: may not specify more than 1 handler type
            tcpSocket:
              port: 5432
            initialDelaySeconds: 0 # default
            successThreshold: 1    # default
            failureThreshold: 3    # default
            periodSeconds: 10      # default (interval)
            timeoutSeconds: 1      # default
          livenessProbe:
            #httpGet:
            #  path: /healthz
            #  port: 8080
            #  #scheme: HTTPS
            tcpSocket:
              port: 5432
            initialDelaySeconds: 0 # default
            successThreshold: 1    # default
            failureThreshold: 3    # default
            periodSeconds: 10      # default (interval)
            timeoutSeconds: 1      # default
          volumeMounts:
            - name: pgdata
              mountPath: /var/lib/pgsql
            - name: env
              mountPath: /app/.env
              subPath: .env
          #lifecycle:
          #  postStart:
          #    exec:
          #      command: ["/bin/sh", "-c", "echo Hello from the postStart handler > /usr/share/message"]
          #  preStop:
          #    exec:
          #      # recent nginx-ingress already has this, but still a decent example to gracefully finish connections on your own nginx via SIGQUIT not SIGTERM
          #      # also remember to increase terminationGracePeriodSeconds
          #      command: ["/bin/sh","-c","nginx -s quit; while killall -0 nginx; do sleep 1; done"]

        #
        # GCP Cloud SQL Proxy sidecar - app can then connect to DB at localhost
        #
        #   https://cloud.google.com/sql/docs/postgres/connect-kubernetes-engine
        #
        # Set up Workload Identity integration between k8s SA + GCP SA
        #
        #   https://cloud.google.com/sql/docs/postgres/connect-kubernetes-engine#workload-identity
        #
        - name: cloud-sql-proxy
          image: gcr.io/cloudsql-docker/gce-proxy:1.17
          command:
            - /cloud_sql_proxy
            - "-ip_address_types=PRIVATE"
            - "-instances=$(CLOUD_SQL_INSTANCE)=tcp:5432"  # populate CLOUD_SQL_INSTANCE env var via configmap, must also specify a port (eg. 5432 for PostgreSQL or 3306 for MySQL, or else a -dir option if using socket dir)
          securityContext:
            # The default Cloud SQL proxy image runs as the
            # "nonroot" user and group (uid: 65532) by default.
            runAsNonRoot: true
            privileged: false
            allowPrivilegeEscalation: false
            readOnlyRootFilesystem: true
            capabilities:
              drop:
                - ALL
          resources:
            requests:
              cpu: 100m
              memory: 128mi
            limits:
              cpu: "1"
              memory: 2gi
          readinessProbe:
            tcpSocket:
              #port: 3306  # mysql
              port: 5432   # postgres
            periodSeconds: 5
          livenessProbe:
            tcpSocket:
              #port: 3306  # mysql
              port: 5432   # postgres
            failureThreshold: 10
      volumes:
        - name: env
          configMap:
            name: myconfigmap
            defaultMode: 0400
            items:
              - key: .env
                path: .env
  volumeClaimTemplates:
    - metadata:
        name: pgdata
      spec:
        # XXX: use a storageclass-*.yaml to:
        #      - set a "reclaimPolicy: Retain" for safety to avoid data loss if deleting/recreating the namespace with the pvc
        #        (it still requires a manual recovery for this easy mistake to make but at least the data isn't deleted)
        #      - choose disk type (SSD vs spinning disk)
        #      - make disk dynamically resizeable (eg. on GCP)
        #storageClassName: ssd
        #storageClassName: ssd-resizeable
        #storageClassName: standard-resizeable
        accessModes:
          - ReadWriteOnce
        resources:
          requests:
            # resize here doesn't affect the PVC, must edit the PVC directly to trigger the provider resize, didn't even need a pod cycle in GCP testing
            storage: 10Gi