services:
  db:
    image: postgres:15
    container_name: rss_db
    environment:
      - POSTGRES_DB=${DB_NAME}
      - POSTGRES_USER=${DB_USER}
      - POSTGRES_PASSWORD=${DB_PASS}
    volumes:
      - postgres_data:/var/lib/postgresql/data
      - ./init-db:/docker-entrypoint-initdb.d
    restart: always
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U ${DB_USER} -d ${DB_NAME}"]
      interval: 5s
      timeout: 5s
      retries: 5

  web:
    build:
      context: .
      args:
        # La imagen llevará torch-cu121 por reutilizar Dockerfile; web no usa GPU.
        TORCH_CUDA: cu121
    container_name: rss_web
    command: gunicorn --bind 0.0.0.0:8000 --workers 3 app:app
    ports:
      - "8001:8000"
    environment:
      - DB_HOST=db
      - DB_PORT=5432
      - DB_NAME=${DB_NAME}
      - DB_USER=${DB_USER}
      - DB_PASS=${DB_PASS}
      - SECRET_KEY=${SECRET_KEY}
      # - NEWS_PER_PAGE=20   # opcional
    depends_on:
      db:
        condition: service_healthy
    restart: always

  scheduler:
    build:
      context: .
      args:
        TORCH_CUDA: cu121
    container_name: rss_scheduler
    command: python scheduler.py
    environment:
      - DB_HOST=db
      - DB_PORT=5432
      - DB_NAME=${DB_NAME}
      - DB_USER=${DB_USER}
      - DB_PASS=${DB_PASS}
      - SECRET_KEY=${SECRET_KEY}
    depends_on:
      db:
        condition: service_healthy
    restart: always

  translator:
    build:
      context: .
      args:
        TORCH_CUDA: cu121  # PyTorch con CUDA 12.1 en la imagen
    container_name: rss_translator
    command: python translation_worker.py
    environment:
      # --- DB ---
      - DB_HOST=db
      - DB_PORT=5432
      - DB_NAME=${DB_NAME}
      - DB_USER=${DB_USER}
      - DB_PASS=${DB_PASS}

      # --- Worker ---
      - TARGET_LANGS=es
      - TRANSLATOR_BATCH=4          # 1.3B: más seguro en 12 GB (sube a 4 si ves VRAM libre)
      - ENQUEUE=200
      - TRANSLATOR_SLEEP_IDLE=5

      # Tokens (equilibrio calidad/VRAM)
      - MAX_SRC_TOKENS=512
      - MAX_NEW_TOKENS=256

      # Beams: mejor título, cuerpo eficiente
      - NUM_BEAMS_TITLE=3
      - NUM_BEAMS_BODY=2

      # Modelo NLLB 1.3B
      - UNIVERSAL_MODEL=facebook/nllb-200-1.3B

      # Dispositivo (forzar GPU si está disponible; el worker cae a CPU si hay OOM)
      - DEVICE=cuda

      # Rendimiento / estabilidad
      - PYTHONUNBUFFERED=1
      - HF_HOME=/root/.cache/huggingface
      - TOKENIZERS_PARALLELISM=false
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:64,garbage_collection_threshold:0.9

      # GPU (requiere NVIDIA Container Toolkit en el host)
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
    volumes:
      - hf_cache:/root/.cache/huggingface
    depends_on:
      db:
        condition: service_healthy
    restart: always
    # Habilita GPU (Compose v2 + nvidia-container-toolkit)
    gpus: all
    # Alternativa con 'deploy':
    # deploy:
    #   resources:
    #     reservations:
    #       devices:
    #         - capabilities: [gpu]

volumes:
  postgres_data:
  hf_cache: