version: "3.8"

services:
  # Production Gradio Demo service
  step-rl-demo:
    build:
      context: .
      dockerfile: Dockerfile
    image: step-rl:latest
    container_name: step-rl-demo-prod
    ports:
      - "7860:7860"
    volumes:
      - ./outputs:/app/outputs
      - ./data:/app/data
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    environment:
      - CUDA_VISIBLE_DEVICES=0
      - PYTHONUNBUFFERED=1
      - HF_HOME=/app/.cache/huggingface
      - WANDB_PROJECT=step-rl
    healthcheck:
      test: ["CMD", "python", "-c", "import torch; print('ok')"]
      interval: 30s
      timeout: 10s
      start_period: 5s
      retries: 3
    restart: unless-stopped

  # Production training service
  step-rl-train:
    build:
      context: .
      dockerfile: Dockerfile
    image: step-rl:latest
    container_name: step-rl-train-prod
    command: python -m step_rl.training.grpo_trainer --config config.yaml
    volumes:
      - ./data:/app/data
      - ./outputs:/app/outputs
      - ./checkpoints:/app/checkpoints
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      - CUDA_VISIBLE_DEVICES=0,1,2,3
      - PYTHONUNBUFFERED=1
      - HF_HOME=/app/.cache/huggingface
      - WANDB_PROJECT=step-rl
      - NCCL_DEBUG=WARN
    healthcheck:
      test: ["CMD", "python", "-c", "import torch; print('ok')"]
      interval: 30s
      timeout: 10s
      start_period: 5s
      retries: 3
    restart: unless-stopped

  # Prometheus metrics scraping target
  step-rl-metrics:
    build:
      context: .
      dockerfile: Dockerfile
    image: step-rl:latest
    container_name: step-rl-metrics-prod
    command: python -m step_rl.monitoring.metrics_server
    ports:
      - "9090:9090"
    environment:
      - PYTHONUNBUFFERED=1
    healthcheck:
      test: ["CMD", "python", "-c", "import prometheus_client; print('ok')"]
      interval: 30s
      timeout: 10s
      start_period: 5s
      retries: 3
    restart: unless-stopped