5. Deploy Ayakaleaf Pro

Deploy the Ayakaleaf Pro pod, its data volume and a NodePort Service.

40-overleaf.yaml

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: overleaf-data
  namespace: overleaf
spec:
  accessModes: [ReadWriteOnce]
  resources:
    requests:
      storage: 20Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: overleaf
  namespace: overleaf
spec:
  replicas: 1
  strategy:
    type: Recreate          # one pod owns the data volume; never run two at once
  selector:
    matchLabels:
      app: overleaf
  template:
    metadata:
      labels:
        app: overleaf
    spec:
      terminationGracePeriodSeconds: 60   # time to flush pending edits, like the Toolkit
      # Keep the environment to what the ConfigMap and Secret set: no REDIS_PORT=tcp://... and the like
      enableServiceLinks: false
      containers:
        - name: sharelatex
          image: ghcr.io/ayaka-notes/overleaf-pro:6.3.2
          envFrom:
            - configMapRef:
                name: overleaf-env
            - secretRef:
                name: overleaf-secrets
          ports:
            - containerPort: 80
          readinessProbe:
            httpGet:
              path: /status
              port: 80
            periodSeconds: 10
          resources:
            requests:
              cpu: 500m
              memory: 1536Mi
            limits:
              memory: 3Gi
          volumeMounts:
            - name: data
              mountPath: /var/lib/overleaf
            - name: compiles
              mountPath: /var/lib/overleaf/data/compiles
            - name: output
              mountPath: /var/lib/overleaf/data/output
            - name: cache
              mountPath: /var/lib/overleaf/data/cache
            - name: docker-sock
              mountPath: /var/run/docker.sock
      volumes:
        - name: data
          persistentVolumeClaim:
            claimName: overleaf-data
        # Compile scratch space on the node; same paths as SANDBOXED_COMPILES_HOST_DIR_*
        - name: compiles
          hostPath:
            path: /srv/overleaf/compiles
            type: DirectoryOrCreate
        - name: output
          hostPath:
            path: /srv/overleaf/output
            type: DirectoryOrCreate
        - name: cache
          hostPath:
            path: /srv/overleaf/cache
            type: DirectoryOrCreate
        - name: docker-sock
          hostPath:
            path: /var/run/docker.sock
            type: Socket
---
apiVersion: v1
kind: Service
metadata:
  name: overleaf
  namespace: overleaf
spec:
  type: NodePort
  selector:
    app: overleaf
  ports:
    - name: http
      port: 80
      targetPort: 80
      nodePort: 30080

kubectl apply -f 40-overleaf.yaml
kubectl -n overleaf rollout status deployment/overleaf --timeout=10m
kubectl -n overleaf logs deployment/overleaf | grep -E 'Finished migrations|Missing required'

The first start runs the database migrations and takes a minute or two. The pod is ready when /status answers; then open OVERLEAF_SITE_URL.

Why the manifest looks like this

• Data volume. Everything under /var/lib/overleaf (project files, history, uploads) lives on the PVC. The data stays on one node with node-local storage such as local-path; with network block storage the pod can move between nodes.
• Compile volumes and the Docker socket. Ayakaleaf Pro asks the node’s Docker to start a TeX Live container and passes it the project’s compile directory by its path on the node. These directories are therefore hostPath volumes, and their paths are repeated in SANDBOXED_COMPILES_HOST_DIR_*. Their content is scratch data and needs no backup.
• enableServiceLinks: false. Kubernetes adds Docker-style variables for every Service of the namespace to each pod: the Service redis becomes REDIS_PORT=tcp://10.x.x.x:6379, the Service git-bridge GIT_BRIDGE_PORT=tcp://.... Some Overleaf services read variables with these names. In this pod the settings from OVERLEAF_REDIS_* and the ConfigMap win, but git-bridge takes GIT_BRIDGE_PORT as its port and stops (see Git integration). Turning the links off keeps the environment to what you set.
• Recreate. Two pods must never use the same data volume or run migrations at the same time.
• Exposing it. The NodePort Service answers on port 30080 of every node. In production put your ingress controller or load balancer in front and terminate TLS there; it must allow WebSockets, request timeouts of 10 minutes (long compiles) and uploads of 50 MB, and pass X-Forwarded-For and X-Forwarded-Proto.