3. Deploy MongoDB and Redis

Run MongoDB as a replica set and Redis with persistence, or point to your own.

Ayakaleaf Pro needs:

• MongoDB 8 as a replica set (it uses transactions). Three members spread over three nodes survive the loss of a node; one member also works.
• Redis 7, one server with AOF persistence. Redis holds the edits that are not yet written to MongoDB, the sessions, and the messages between Overleaf services. Overleaf connects to a single Redis server: Sentinel and Redis Cluster are not supported.

You can use managed services instead, such as MongoDB Atlas or your cloud’s Redis. Skip this page and set OVERLEAF_MONGO_URL and OVERLEAF_REDIS_HOST to them in step 4.

1. MongoDB

   10-mongo.yaml

   apiVersion: v1
   kind: Service
   metadata:
     name: mongo
     namespace: overleaf
   spec:
     clusterIP: None                 # headless: gives mongo-0.mongo, mongo-1.mongo, ...
     publishNotReadyAddresses: true  # members must resolve each other before the set is initiated
     selector:
       app: mongo
     ports:
       - name: mongo
         port: 27017
   ---
   apiVersion: apps/v1
   kind: StatefulSet
   metadata:
     name: mongo
     namespace: overleaf
   spec:
     serviceName: mongo
     replicas: 3
     selector:
       matchLabels:
         app: mongo
     template:
       metadata:
         labels:
           app: mongo
       spec:
         affinity:
           podAntiAffinity:
             requiredDuringSchedulingIgnoredDuringExecution:
               - labelSelector:
                   matchLabels:
                     app: mongo
                 topologyKey: kubernetes.io/hostname
         containers:
           - name: mongo
             image: mongo:8.0
             args:
               - --replSet=overleaf
               - --bind_ip_all
               - --wiredTigerCacheSizeGB=0.25   # default is 50% of (RAM - 1 GB); size it for your nodes
             # env:
             #   - name: GLIBC_TUNABLES        # only on kernel 6.19+ (SERVER-121912)
             #     value: glibc.pthread.rseq=1
             ports:
               - containerPort: 27017
             readinessProbe:
               tcpSocket:
                 port: 27017
             resources:
               requests:
                 cpu: 100m
                 memory: 384Mi
               limits:
                 memory: 768Mi
             volumeMounts:
               - name: data
                 mountPath: /data/db
     volumeClaimTemplates:
       - metadata:
           name: data
         spec:
           accessModes: [ReadWriteOnce]
           resources:
             requests:
               storage: 5Gi
   ---
   # Initiates the replica set once. Safe to re-run: it does nothing when the set exists.
   apiVersion: batch/v1
   kind: Job
   metadata:
     name: mongo-init-replica-set
     namespace: overleaf
   spec:
     backoffLimit: 20
     template:
       spec:
         restartPolicy: OnFailure
         containers:
           - name: init
             image: mongo:8.0
             command:
               - mongosh
               - --quiet
               - --host=mongo-0.mongo
               - --eval
               - |
                 try {
                   rs.status()
                   print('replica set already initiated')
                 } catch (e) {
                   rs.initiate({
                     _id: 'overleaf',
                     members: [
                       { _id: 0, host: 'mongo-0.mongo.overleaf.svc.cluster.local:27017' },
                       { _id: 1, host: 'mongo-1.mongo.overleaf.svc.cluster.local:27017' },
                       { _id: 2, host: 'mongo-2.mongo.overleaf.svc.cluster.local:27017' },
                     ],
                   })
                 }
   

   kubectl apply -f 10-mongo.yaml
   kubectl -n overleaf rollout status statefulset/mongo
   kubectl -n overleaf wait --for=condition=complete job/mongo-init-replica-set --timeout=5m
   kubectl -n overleaf exec mongo-0 -- mongosh --quiet --eval 'rs.status().members.map(m => m.name + " " + m.stateStr)'
   

   One member should be PRIMARY and the others SECONDARY after a few seconds. The Job retries until all members answer, so a few failed Job pods are normal.

   • The headless Service needs publishNotReadyAddresses: true: the members must find each other by name before the replica set exists.
   • The members are added with their stable names (mongo-0.mongo.overleaf.svc.cluster.local), never pod IPs, which change on every restart.
   • For a single member, set replicas: 1 and keep only mongo-0 in the Job.
2. Redis

   20-redis.yaml

   apiVersion: v1
   kind: Service
   metadata:
     name: redis
     namespace: overleaf
   spec:
     selector:
       app: redis
     ports:
       - name: redis
         port: 6379
   ---
   apiVersion: apps/v1
   kind: StatefulSet
   metadata:
     name: redis
     namespace: overleaf
   spec:
     serviceName: redis
     replicas: 1
     selector:
       matchLabels:
         app: redis
     template:
       metadata:
         labels:
           app: redis
       spec:
         containers:
           - name: redis
             image: redis:7.4
             args: [--appendonly, "yes"]
             ports:
               - containerPort: 6379
             readinessProbe:
               tcpSocket:
                 port: 6379
             resources:
               requests:
                 cpu: 50m
                 memory: 128Mi
               limits:
                 memory: 512Mi
             volumeMounts:
               - name: data
                 mountPath: /data
     volumeClaimTemplates:
       - metadata:
           name: data
         spec:
           accessModes: [ReadWriteOnce]
           resources:
             requests:
               storage: 2Gi
   

   kubectl apply -f 20-redis.yaml
   kubectl -n overleaf rollout status statefulset/redis
   

Q&A

MongoDB 8 crashes at startup on a recent kernel. Why?

Answer: MongoDB 8 fails on Linux kernel 6.19 and later (SERVER-121912). Uncomment the GLIBC_TUNABLES environment variable in 10-mongo.yaml.

How much memory does MongoDB need?

Answer: By default MongoDB uses half of the node’s memory minus 1 GB for its cache, and it sees the node’s memory, not the pod’s limit. Set --wiredTigerCacheSizeGB to fit your nodes; 0.25 kept each member at about 200 to 260 MiB on our 4 GB test nodes.