水平扩展

在使用会话保持的负载均衡器之后运行多个 Ayakaleaf Pro 副本。

一个副本的每次编译都只使用一个节点的 CPU：在我们配备 2 个 vCPU 的测试节点上，六个同时进行的编译所花的时间是单个编译的三倍，而其他节点却处于空闲状态。多个副本可以将用户及其编译分散到各个节点上。

本方案遵循 Overleaf 的水平扩展设计，并将其移植到 Kubernetes：

要求	本方案中的实现
共享的 MongoDB 和 Redis	安装时的 10-mongo.yaml 和 20-redis.yaml
共享的密钥	第 2 步中的 Secret overleaf-secrets 已包含它们
项目文件和历史记录存放在 S3 中（而非 NFS）	任意 S3 服务，参见 S3；该页以 Garage 作为集群内示例
每个副本一块本地磁盘，用于编译和上传	每个节点上的 hostPath /srv/overleaf-ha
持久路由、WebSocket、10 分钟超时、50 MB 上传	使用 Overleaf 示例配置的 HAProxy
git-bridge 作为单例，与一个副本相邻	以 overleaf-0 作为其相邻副本的 git-bridge

请从一开始就规划好 S3。已在其卷上保存文件的实例需要先进行 S3 迁移。本方案将替代第 5 步中的 Deployment：在应用 StatefulSet 之前请先删除它（kubectl delete -f 40-overleaf.yaml）。

1. S3 存储

   按照 S3 中的说明，使用你的 S3 服务或集群中的 Garage 设置三个存储桶和两个用户，并将四个密钥放入 Secret overleaf-s3。下面的配置使用该页中的 Garage 示例。

2. 配置

   此 ConfigMap 替代 30-config.yaml。它添加了 S3 和 git-bridge，并将 OVERLEAF_SITE_URL 指向负载均衡器：

   60-config.yaml

   apiVersion: v1
   kind: ConfigMap
   metadata:
     name: overleaf-env
     namespace: overleaf
   data:
     OVERLEAF_APP_NAME: Ayakaleaf Pro
     OVERLEAF_SITE_URL: http://192.0.2.10:30081      # the load balancer (HAProxy) URL
     OVERLEAF_ADMIN_EMAIL: [e-mail]
     # Emails: add the sender and the SMTP server together, see Email delivery
     OVERLEAF_MONGO_URL: mongodb://mongo-0.mongo:27017,mongo-1.mongo:27017,mongo-2.mongo:27017/sharelatex?replicaSet=overleaf
     OVERLEAF_REDIS_HOST: redis
     REDIS_HOST: redis
     ENABLED_LINKED_FILE_TYPES: project_file,project_output_file,url
     ENABLE_CONVERSIONS: "true"
     EMAIL_CONFIRMATION_DISABLED: "true"
     OVERLEAF_TRUSTED_PROXY_IPS: loopback,10.244.0.0/16   # HAProxy runs in the pod CIDR
     # Sandboxed compiles, node paths: each replica owns /srv/overleaf-ha on its node
     SANDBOXED_COMPILES: "true"
     SANDBOXED_COMPILES_SIBLING_CONTAINERS: "true"
     DOCKER_RUNNER: "true"
     SANDBOXED_COMPILES_HOST_DIR_COMPILES: /srv/overleaf-ha/data/compiles
     SANDBOXED_COMPILES_HOST_DIR_OUTPUT: /srv/overleaf-ha/data/output
     SANDBOXED_COMPILES_HOST_DIR_CACHE: /srv/overleaf-ha/data/cache
     ALL_TEX_LIVE_DOCKER_IMAGES: ghcr.io/ayaka-notes/texlive-full:2026.1
     ALL_TEX_LIVE_DOCKER_IMAGE_NAMES: TeX Live 2026
     TEX_LIVE_DOCKER_IMAGE: ghcr.io/ayaka-notes/texlive-full:2026.1
     # S3 (required for more than one replica); keys are in the Secret overleaf-s3
     OVERLEAF_FILESTORE_BACKEND: s3
     OVERLEAF_FILESTORE_TEMPLATE_FILES_BUCKET_NAME: overleaf-template-files
     OVERLEAF_FILESTORE_S3_ENDPOINT: http://garage:3900
     OVERLEAF_FILESTORE_S3_PATH_STYLE: "true"
     OVERLEAF_FILESTORE_S3_REGION: garage
     OVERLEAF_HISTORY_BACKEND: s3
     OVERLEAF_HISTORY_PROJECT_BLOBS_BUCKET: overleaf-project-blobs
     OVERLEAF_HISTORY_CHUNKS_BUCKET: overleaf-chunks
     OVERLEAF_HISTORY_S3_ENDPOINT: http://garage:3900
     OVERLEAF_HISTORY_S3_PATH_STYLE: "true"
     OVERLEAF_HISTORY_S3_REGION: garage
     # Git integration: one git-bridge, its sibling is overleaf-0
     GIT_BRIDGE_ENABLED: "true"
     GIT_BRIDGE_HOST: 10.96.100.80      # the git-bridge Service's fixed ClusterIP, see 75-git-bridge.yaml
     GIT_BRIDGE_PORT: "8000"
   
3. 副本

   70-overleaf.yaml

   apiVersion: v1
   kind: Service
   metadata:
     name: overleaf-hl
     namespace: overleaf
   spec:
     clusterIP: None            # stable DNS per replica: overleaf-0.overleaf-hl, ...
     selector:
       app: overleaf-ha
     ports:
       - {name: http, port: 80}
       - {name: web, port: 3000}       # git-bridge API (sibling only)
       - {name: history, port: 3100}   # history-v1, used by git-bridge via V1_HISTORY_URL
   ---
   apiVersion: apps/v1
   kind: StatefulSet
   metadata:
     name: overleaf
     namespace: overleaf
   spec:
     serviceName: overleaf-hl
     replicas: 3
     # OrderedReady: overleaf-0 runs the migrations alone before the next replica starts
     podManagementPolicy: OrderedReady
     selector:
       matchLabels:
         app: overleaf-ha
     template:
       metadata:
         labels:
           app: overleaf-ha
       spec:
         terminationGracePeriodSeconds: 60
         # Keep the environment to the ConfigMap and Secrets: no REDIS_PORT=tcp://..., GIT_BRIDGE_PORT=tcp://...
         enableServiceLinks: false
         affinity:
           podAntiAffinity:            # one replica per node: each uses that node's Docker and disk
             requiredDuringSchedulingIgnoredDuringExecution:
               - labelSelector:
                   matchLabels:
                     app: overleaf-ha
                 topologyKey: kubernetes.io/hostname
         containers:
           - name: sharelatex
             image: ghcr.io/ayaka-notes/overleaf-pro:6.3.2
             env:
               - name: POD_NAME
                 valueFrom:
                   fieldRef:
                     fieldPath: metadata.name
               # Each replica hands out history URLs with its own stable name, so git-bridge can reach them
               - name: V1_HISTORY_URL
                 value: http://$(POD_NAME).overleaf-hl.overleaf.svc.cluster.local:3100/api
             envFrom:
               - configMapRef: {name: overleaf-env}
               - secretRef: {name: overleaf-secrets}
               - secretRef: {name: overleaf-s3}
             ports:
               - containerPort: 80
             readinessProbe:
               httpGet: {path: /status, port: 80}
               periodSeconds: 10
             resources:
               requests: {cpu: 500m, memory: 1536Mi}
               limits: {memory: 3Gi}
             volumeMounts:
               - {name: local, mountPath: /var/lib/overleaf}
               - {name: docker-sock, mountPath: /var/run/docker.sock}
         volumes:
           # Local disk for ephemeral files (compiles, output, uploads); the HOST_DIR_* paths are under it
           - name: local
             hostPath: {path: /srv/overleaf-ha, type: DirectoryOrCreate}
           - name: docker-sock
             hostPath: {path: /var/run/docker.sock, type: Socket}
   

   • 使用 StatefulSet，而非 Deployment。每个副本都会获得一个稳定的名称（overleaf-0、overleaf-1……），负载均衡器 Cookie 和 git-bridge 都引用该名称；OrderedReady 会先单独启动 overleaf-0，因此只有一个副本运行数据库迁移。
   • 每个节点一个副本。每个副本都通过其所在节点的 Docker 启动编译，并将编译内容保存在该节点的 /srv/overleaf-ha 中。
   • 每个副本各自的 V1_HISTORY_URL。每个副本都使用自己的名称分发历史记录链接，以便 git-bridge 能够访问它们。
4. git-bridge

   75-git-bridge.yaml

   apiVersion: v1
   kind: Service
   metadata:
     name: git-bridge
     namespace: overleaf
   spec:
     # Fixed ClusterIP (inside the service CIDR): the sharelatex image's nginx proxies /git/ with
     # `resolver 127.0.0.11` (Docker's DNS), which does not exist in Kubernetes, so GIT_BRIDGE_HOST is this IP
     clusterIP: 10.96.100.80
     selector:
       app: git-bridge
     ports:
       - {name: http, port: 8000}
   ---
   apiVersion: apps/v1
   kind: StatefulSet
   metadata:
     name: git-bridge
     namespace: overleaf
   spec:
     serviceName: git-bridge
     replicas: 1
     selector:
       matchLabels:
         app: git-bridge
     template:
       metadata:
         labels:
           app: git-bridge
       spec:
         # Service "git-bridge" would inject GIT_BRIDGE_PORT=tcp://..., which git-bridge reads as its port
         enableServiceLinks: false
         containers:
           - name: git-bridge
             image: ghcr.io/ayaka-notes/overleaf-pro/git-bridge:6.3.2   # same version as the sharelatex image
             command: ["/server-pro-start.sh"]
             securityContext:
               runAsUser: 0
             env:
               - {name: GIT_BRIDGE_API_BASE_URL, value: "http://overleaf-0.overleaf-hl.overleaf.svc.cluster.local:3000/api/v0/"}
               - {name: GIT_BRIDGE_OAUTH2_SERVER, value: "http://overleaf-0.overleaf-hl.overleaf.svc.cluster.local"}
               - {name: GIT_BRIDGE_POSTBACK_BASE_URL, value: "http://git-bridge.overleaf.svc.cluster.local:8000"}
               - {name: GIT_BRIDGE_ROOT_DIR, value: /data/git-bridge}
             ports:
               - containerPort: 8000
             resources:
               requests: {cpu: 100m, memory: 256Mi}
               limits: {memory: 768Mi}
             volumeMounts:
               - {name: data, mountPath: /data/git-bridge}
     volumeClaimTemplates:
       - metadata:
           name: data
         spec:
           accessModes: [ReadWriteOnce]
           resources:
             requests:
               storage: 5Gi
   

   关于 GIT_BRIDGE_HOST 为什么是固定 IP 地址，请参见 Git 集成。

5. 负载均衡器

   该 HAProxy 配置基于 Overleaf 的示例，并做了五处修改：/project/<id>/mcp 与 /git/ 一样路由到 overleaf-0（参见 MCP 服务器）；在端口 8080 上使用普通 HTTP（该镜像以非 root 用户运行，在 containerd 下无法绑定端口 80）；通过集群 DNS 查找服务器（Pod IP 在每次重启时都会变化）；使用 DaemonSet，使每个节点都能响应；以及使用 balance leastconn 代替默认的 roundrobin（参见下面的 Q&A）。

   80-haproxy.yaml

   apiVersion: v1
   kind: ConfigMap
   metadata:
     name: haproxy-config
     namespace: overleaf
   data:
     haproxy.cfg: |
       global
         log stdout format raw local0 info
   
       resolvers kubedns
         parse-resolv-conf
         hold valid 10s
   
       defaults
         mode                    http
         # New sessions go to the replica with the fewest open connections (editors keep a WebSocket open)
         balance                 leastconn
         option                  httpchk HEAD /status
         http-check              expect status 200
         default-server          check resolvers kubedns init-addr none
         log                     global
         option                  httplog
         # Reroute to a different backend if the sticky one is down
         option                  redispatch 1
         retries                 3
         # Sticky session for 24h of inactivity -- compile output is deleted after 24h
         cookie                  server-pro-ha insert maxidle 24h
         timeout queue           1m
         timeout connect         15s
         timeout client          1m
         # Allow slow compiles -- hard-coded limit in clsi is 10min
         timeout server          10m
         # Disconnect the editor after 23h
         timeout tunnel          23h
         # Lower than NGINX_KEEPALIVE_TIMEOUT (65s) in the sharelatex container
         timeout http-keep-alive 60s
   
       listen overleaf
         bind :8080
         option forwardfor
         # When TLS ends in front of HAProxy, also pass:  http-request set-header X-Forwarded-Proto https
         # git requests go to the git-bridge sibling
         use-server overleaf-0 if { path_beg /git/ }
         # MCP clients send no cookie: keep compile, log and PDF download on one replica
         use-server overleaf-0 if { path_reg ^/project/[0-9a-f]{24}/mcp(/|$) }
         http-response add-header X-Served-By %s
         stats enable
         stats uri /haproxy
         server overleaf-0 overleaf-0.overleaf-hl.overleaf.svc.cluster.local:80 cookie overleaf-0
         server overleaf-1 overleaf-1.overleaf-hl.overleaf.svc.cluster.local:80 cookie overleaf-1
         server overleaf-2 overleaf-2.overleaf-hl.overleaf.svc.cluster.local:80 cookie overleaf-2
   ---
   apiVersion: apps/v1
   # DaemonSet: one HAProxy per node, so the NodePort answers on every node with externalTrafficPolicy
   # Local. The cookie holds the replica name, so stickiness works whichever HAProxy a request hits.
   kind: DaemonSet
   metadata:
     name: haproxy
     namespace: overleaf
   spec:
     selector:
       matchLabels:
         app: haproxy
     template:
       metadata:
         labels:
           app: haproxy
         annotations:
           config-hash: "1"   # bump to restart after editing the ConfigMap
       spec:
         containers:
           - name: haproxy
             image: haproxy:lts-alpine
             ports:
               - containerPort: 8080
             resources:
               requests: {cpu: 50m, memory: 32Mi}
               limits: {memory: 256Mi}
             volumeMounts:
               - {name: config, mountPath: /usr/local/etc/haproxy}
         volumes:
           - name: config
             configMap:
               name: haproxy-config
   ---
   apiVersion: v1
   kind: Service
   metadata:
     name: haproxy
     namespace: overleaf
   spec:
     type: NodePort
     externalTrafficPolicy: Local    # keep the client IP (no SNAT); only nodes running HAProxy answer
     selector:
       app: haproxy
     ports:
       - {name: http, port: 80, targetPort: 8080, nodePort: 30081}
   

   kubectl delete -f 40-overleaf.yaml --ignore-not-found
   kubectl apply -f 60-config.yaml -f 70-overleaf.yaml -f 75-git-bridge.yaml -f 80-haproxy.yaml
   kubectl -n overleaf rollout status statefulset/overleaf --timeout=20m
   curl -sI http://192.0.2.10:30081/login | grep -i x-served-by
   

   externalTrafficPolicy: Local 会在 X-Forwarded-For 中保留用户的 IP 地址。

6. TLS

   按照 TLS 代理中的说明在 HAProxy 中终止 HTTPS，然后在 60-config.yaml 中将 OVERLEAF_SITE_URL 设置为 https:// 地址，并设置 OVERLEAF_SECURE_COOKIE: "true"。

​

预期效果

我们在三个各配备 2 个 vCPU 和 4 GB 内存的节点上测试了三个副本，MongoDB、Redis、Garage、git-bridge 和 HAProxy 也运行在这些节点上：

• 12 名用户同时编译一篇 600 页的文档：全部 12 名用户都完成了编译并获取到 PDF，每个副本 4 名用户；一轮耗时 8 到 18 秒。每个节点至少保留了 1.3 GB 可用内存。
• 每个用户都停留在同一个副本上（Cookie server-pro-ha），包括必须来自执行编译的副本的 PDF 下载。在 HTTPS 下同样如此，包括编辑器的 WebSocket。
• 当某个副本停止时，其用户会在大约 15 到 20 秒内遇到错误，直到 HAProxy 将其标记为下线；之后用户会转移到另一个副本并保持登录状态，因为会话保存在 Redis 中。

​

Q&A

为什么使用 balance leastconn？

回答：在我们的测试中，使用 Overleaf 的示例配置（roundrobin，每个节点一个 HAProxy）时，12 名新用户在三个副本上的分布为 8、3 和 1，最繁忙的副本让最慢的用户等待了 29 秒。使用 leastconn（优先选择打开连接数最少的副本，每个编辑器都会保持一个 WebSocket 连接）时，同样的测试得到的分布为 4、4 和 4，耗时 18 秒。会话保持不受影响：它来自 Cookie。

我可以使用自己的 Ingress 控制器代替 HAProxy 吗？

回答：可以，前提是它能根据 Cookie 路由到固定的副本、支持 WebSocket、允许 10 分钟的请求时间和 50 MB 的上传、将其 keep-alive 超时保持在 65 秒以下、发送 X-Forwarded-For 和 X-Forwarded-Proto，并将 /git/ 和 /project/<id>/mcp 发送到 overleaf-0。请路由到各副本自身的名称（overleaf-N.overleaf-hl），而不是路由到对每个请求进行负载均衡的 Service。

如何添加第四个副本？

回答：添加一个安装了 Docker 和 TeX Live 镜像的节点，在 70-overleaf.yaml 中设置 replicas: 4，并在 HAProxy 配置中添加一行 server overleaf-3 ...（同时递增 config-hash，以便 HAProxy 重启）。