Foundry/swarm/docker/docker-stack.yml

# Foundry Swarm - Docker Stack for production Swarm deployment
#
# Prerequisites:
#   1. Initialize swarm: docker swarm init
#   2. Join other nodes: docker swarm join --token <token> <manager-ip>:2377
#   3. Label nodes by capability:
#      docker node update --label-add hw=server <server-node>
#      docker node update --label-add hw=desktop <desktop-node>
#      docker node update --label-add hw=laptop <laptop-node>
#      docker node update --label-add hw=rpi <rpi-node>
#
# Deployment:
#   docker stack deploy -c docker-stack.yml foundry
#
# Scaling:
#   docker service scale foundry_worker-desktop=5
#   docker service scale foundry_worker-rpi=10
#
# Monitoring:
#   docker service logs -f foundry_coordinator
#   docker service ps foundry_worker-rpi

version: "3.8"

services:
  # ============================================================
  # COORDINATOR - Runs on the most powerful server
  # ============================================================
  coordinator:
    image: foundry/swarm-coordinator:latest
    ports:
      - target: 7100
        published: 7100
        protocol: tcp
        mode: host  # Host mode for better performance
    volumes:
      - checkpoints:/app/checkpoints
      - logs:/app/logs
      - type: bind
        source: ./config
        target: /app/config
        read_only: true
    environment:
      - SW_PORT=7100
      - SW_MIN_WORKERS=10
      - SW_LEARNING_RATE=1e-4
      - SW_TOTAL_STEPS=500000
      - SW_CHECKPOINT_EVERY=5000
      - SW_HEARTBEAT_TIMEOUT=120    # Longer timeout for WiFi latency
      - SW_SYNC_EVERY_N_GRADS=5     # Don't wait for all workers
    networks:
      - swarm-overlay
    deploy:
      mode: replicated
      replicas: 1
      placement:
        constraints:
          - node.labels.hw == server
      resources:
        limits:
          cpus: "8"
          memory: 32G
        reservations:
          cpus: "2"
          memory: 8G
      restart_policy:
        condition: on-failure
        delay: 5s
        max_attempts: 3

  # ============================================================
  # WORKERS - Different configs per hardware class
  # ============================================================

  # Server workers - highest throughput
  worker-server:
    image: foundry/swarm-worker:latest
    volumes:
      - training-data:/app/data:ro
      - logs:/app/logs
    environment:
      - SW_COORDINATOR_HOST=coordinator
      - SW_COORDINATOR_PORT=7100
      - SW_DATA_PATH=/app/data
      - SW_BATCH_SIZE=256
      - SW_NUM_THREADS=0
      - SW_COMPRESSION=topk
      - SW_TOPK_RATIO=0.1
    networks:
      - swarm-overlay
    deploy:
      mode: replicated
      replicas: 2
      placement:
        constraints:
          - node.labels.hw == server
      resources:
        limits:
          cpus: "16"
          memory: 64G
        reservations:
          cpus: "4"
          memory: 8G
      restart_policy:
        condition: on-failure

  # Desktop workers - medium throughput
  worker-desktop:
    image: foundry/swarm-worker:latest
    volumes:
      - training-data:/app/data:ro
      - logs:/app/logs
    environment:
      - SW_COORDINATOR_HOST=coordinator
      - SW_COORDINATOR_PORT=7100
      - SW_DATA_PATH=/app/data
      - SW_BATCH_SIZE=64
      - SW_NUM_THREADS=0
      - SW_COMPRESSION=topk
      - SW_TOPK_RATIO=0.1
    networks:
      - swarm-overlay
    deploy:
      mode: replicated
      replicas: 10
      placement:
        constraints:
          - node.labels.hw == desktop
      resources:
        limits:
          cpus: "4"
          memory: 16G
        reservations:
          cpus: "1"
          memory: 2G
      restart_policy:
        condition: on-failure

  # Laptop workers - lower throughput, battery-aware
  worker-laptop:
    image: foundry/swarm-worker:latest
    volumes:
      - training-data:/app/data:ro
      - logs:/app/logs
    environment:
      - SW_COORDINATOR_HOST=coordinator
      - SW_COORDINATOR_PORT=7100
      - SW_DATA_PATH=/app/data
      - SW_BATCH_SIZE=32
      - SW_NUM_THREADS=2            # Limit threads for battery
      - SW_COMPRESSION=topk
      - SW_TOPK_RATIO=0.1
    networks:
      - swarm-overlay
    deploy:
      mode: replicated
      replicas: 5
      placement:
        constraints:
          - node.labels.hw == laptop
      resources:
        limits:
          cpus: "2"
          memory: 8G
        reservations:
          cpus: "0.5"
          memory: 1G
      restart_policy:
        condition: on-failure

  # Raspberry Pi workers - low throughput, many nodes
  worker-rpi:
    image: foundry/swarm-worker:arm64
    volumes:
      - training-data:/app/data:ro
      - logs:/app/logs
    environment:
      - SW_COORDINATOR_HOST=coordinator
      - SW_COORDINATOR_PORT=7100
      - SW_DATA_PATH=/app/data
      - SW_BATCH_SIZE=8
      - SW_NUM_THREADS=4
      - SW_COMPRESSION=topk
      - SW_TOPK_RATIO=0.05          # More aggressive compression
    networks:
      - swarm-overlay
    deploy:
      mode: replicated
      replicas: 20
      placement:
        constraints:
          - node.labels.hw == rpi
      resources:
        limits:
          cpus: "4"
          memory: 3G
        reservations:
          cpus: "1"
          memory: 512M
      restart_policy:
        condition: on-failure
        delay: 10s

networks:
  swarm-overlay:
    driver: overlay
    attachable: true

volumes:
  checkpoints:
    driver: local
  logs:
    driver: local
  training-data:
    driver: local
    driver_opts:
      type: nfs
      o: addr=${NFS_SERVER},rw,nolock
      device: ":${NFS_PATH}"