Environment_Infrastructure/roadmap/prod-env/01-swarm-init-multinode.md

01 — Docker Swarm Init (Prod — Multi-Node)

Context

• Repo: iklim.co root
• Environment: prod
• Topology:
  • 3 × app nodes (iklim-app-01/02/03) — all act as Swarm managers AND app workers with type=service label (Raft quorum: 1 can fail)
  • 3 × DB nodes (iklim-db-01/02/03) — join Swarm as workers with role=db label; DB services are placed exclusively on them
• Sizing: app nodes are cpx42, DB nodes are cpx32; see ../../hetzner-sizing-report.md
• All 6 nodes are in the same private network.
• Pipeline trigger: push to prod-env branch → Gitea runner on prod-runner (first app node).
• App Swarm managers: 3 nodes all manager-eligible and carry app workloads with type=service label (no dedicated worker-only app nodes).

Node labeling plan

Node	Role	Swarm role	Labels
iklim-app-01	API services, SWAG, Vault	Manager + Worker	type=service
iklim-app-02	API services replicas	Manager + Worker	type=service
iklim-app-03	API services replicas	Manager + Worker	type=service
iklim-db-01	MongoDB replica + PostgreSQL (Patroni), etcd	Worker	role=db, db-index=01
iklim-db-02	MongoDB replica + PostgreSQL (Patroni), etcd	Worker	role=db, db-index=02
iklim-db-03	MongoDB replica + PostgreSQL (Patroni), etcd	Worker	role=db, db-index=03

Label scheme rationale

App nodes carry type=service, DB nodes carry role=db. The two different label keys are not an inconsistency — they operate on different semantic planes:

• type=service — "this node carries service workload"; determines which node group microservices and infrastructure services (APISIX, Vault, RabbitMQ, Redis, SWAG, etc.) are scheduled on.
• role=db — "this node is a database node"; pins PostgreSQL (Patroni), MongoDB, and their proxy services exclusively to DB nodes.

Docker Swarm's built-in node.role property (manager / worker) does not conflict with the custom node.labels.role label — the placement constraint syntax distinguishes them explicitly:

node.role == manager           ← Swarm built-in (manager/worker distinction)
node.labels.type == service    ← custom label (app node workload target)
node.labels.role == db         ← custom label (DB node workload target)

This scheme is applied consistently across docker-stack-infra.yml and all 10 microservice docker-stack-service.yml files. The test environment uses the same type=service label on its single node, so both environments share the same constraint syntax.

node.role == worker is intentionally not used anywhere. DB nodes are Swarm workers, but targeting them via node.role == worker would also match any future worker-only app nodes. The explicit node.labels.role == db label provides precise, unambiguous targeting regardless of Swarm role.

Step 1 — Init Swarm on iklim-app-01 (the prod-runner node)

MANAGER_IP=$(hostname -I | awk '{print $1}')
if ! docker info --format '{{.Swarm.LocalNodeState}}' | grep -q "active"; then
  docker swarm init --advertise-addr "$MANAGER_IP"
  echo "Swarm initialized on $MANAGER_IP"
else
  echo "Swarm already active"
fi

Step 2 — Get manager join token

docker swarm join-token manager  # for iklim-app-02, iklim-app-03

Save this token — needed on iklim-app-02 and iklim-app-03.

Step 3 — Join iklim-app-02 and iklim-app-03 as managers

SSH into iklim-app-02 and iklim-app-03, run:

docker swarm join --token <MANAGER_TOKEN> 10.20.10.11:2377

Step 4 — Label app nodes

On iklim-app-01, after iklim-app-02 and iklim-app-03 have joined:

for node in iklim-app-01 iklim-app-02 iklim-app-03; do
  docker node update --label-add type=service "$node"
done

Step 5 — Join DB nodes as Swarm workers

Get the worker join token on iklim-app-01:

docker swarm join-token worker

SSH into each DB node and join:

docker swarm join --token <WORKER_TOKEN> 10.20.10.11:2377

Then label them on iklim-app-01:

docker node update --label-add role=db --label-add db-index=01 iklim-db-01
docker node update --label-add role=db --label-add db-index=02 iklim-db-02
docker node update --label-add role=db --label-add db-index=03 iklim-db-03

DB nodes are Swarm workers only — they never become managers. DB services are pinned to them via node.labels.role == db placement constraint. See 08-prod-db-cluster-kurulum.md for DB stack deployment.

Step 6 — Verify

docker node ls

Expected: 6 nodes — 3 with MANAGER STATUS = Leader or Reachable, 3 workers with Ready.

docker node inspect iklim-app-01 --format '{{.Spec.Labels}}'
docker node inspect iklim-db-01 --format '{{.Spec.Labels}}'

Expected: map[type:service] for app nodes, map[db-index:01 role:db] (vb.) for DB nodes.

Step 7 — Confirm init/swarm-init.sh multi-node awareness

The script is idempotent (skips init if already active). Verify:

grep -n "swarm init\|swarm join" init/swarm-init.sh

The prod pipeline runs on iklim-app-01 only. iklim-app-02/03 are joined via Ansible (swarm role), not via the Gitea pipeline.

Placement constraints used in docker-stack-infra.yml

Constraint	Resolves to	Services
node.hostname == iklim-app-01	iklim-app-01 only	SWAG, cert-reloader
node.labels.type == service	iklim-app-01, iklim-app-02, iklim-app-03	Vault, Redis, RabbitMQ, APISIX, Prometheus, Grafana, etcd (idle in prod — APISIX uses Patroni etcd)
node.labels.role == db	iklim-db-01, iklim-db-02, iklim-db-03	PostgreSQL (Patroni), MongoDB, etcd (via docker-stack-db.prod.yml)

SWAG and cert-reloader are pinned to iklim-app-01 (the Floating IP node) because SWAG does not support clustering and must match the public entry point. Vault floats across all service nodes; its TLS cert is read from StorageBox (/mnt/storagebox/ssl) so it is available on whichever node Vault is scheduled on. Microservices carry no placement constraint and are distributed by the Swarm scheduler across all app nodes. DB services are pinned to DB nodes via separate DB stacks.