# PRism services.
#
#   docker compose up --build            # api + worker + redis + mlflow
#   docker compose --profile serving up  # also start vLLM with the LoRA adapter
#
# All secrets come from .env / mounted files - nothing secret is baked into
# images or committed to the repo.

services:
  redis:
    image: redis:7-alpine
    restart: unless-stopped
    volumes:
      - redis-data:/data
    healthcheck:
      test: ["CMD", "redis-cli", "ping"]
      interval: 5s
      timeout: 3s
      retries: 5

  api:
    build: .
    restart: unless-stopped
    ports:
      - "8000:8000"
    env_file:
      - .env
    environment:
      REDIS_URL: redis://redis:6379/0
      # NOTE: with `--profile serving`, point the app at the vLLM service by
      # setting VLLM_BASE_URL=http://vllm:8000 in your .env.
    depends_on:
      redis:
        condition: service_healthy
    volumes:
      # mount the private key read-only; never bake it into the image
      - ./private-key.pem:/app/keys/private-key.pem:ro
      - ./config.yaml:/app/config.yaml:ro
      - feedback-data:/app/feedback_data
    healthcheck:
      test:
        - CMD
        - python
        - -c
        - "import sys,urllib.request; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/healthz', timeout=5).getcode() == 200 else 1)"
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 20s

  worker:
    build: .
    restart: unless-stopped
    env_file:
      - .env
    environment:
      REDIS_URL: redis://redis:6379/0
      # NOTE: with `--profile serving`, set VLLM_BASE_URL=http://vllm:8000 in .env.
    depends_on:
      redis:
        condition: service_healthy
    command: ["python", "-m", "prism.worker"]
    volumes:
      - ./private-key.pem:/app/keys/private-key.pem:ro
      - ./config.yaml:/app/config.yaml:ro
      - feedback-data:/app/feedback_data

  mlflow:
    # Tracking server for training/eval runs (sqlite backend store by default;
    # swap MLFLOW_BACKEND_STORE_URI for Postgres in real deployments).
    # Pinned digest for mlflow v2.16.0, refreshed 2026-10-07.
    image: ghcr.io/mlflow/mlflow:v2.16.0@sha256:c9a14c4341220855b4cceecd55f5b1f7db336ed44bfb30c711a35ddfafe38df2
    restart: unless-stopped
    command: >
      mlflow server
      --backend-store-uri sqlite:////mlflow/mlflow.db
      --host 0.0.0.0 --port 5000
    ports:
      - "5000:5000"
    environment:
      MLFLOW_BACKEND_STORE_URI: ${MLFLOW_BACKEND_STORE_URI:-sqlite:////mlflow/mlflow.db}
    volumes:
      - mlflow-data:/mlflow
    healthcheck:
      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:5000/', timeout=5)"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 30s

  # --- Model serving (profile: serving) ---------------------------------
  # vLLM serving the fine-tuned LoRA adapter over the base model.
  # Needs a GPU host (nvidia-container-toolkit). Start with:
  #   VLLM_BASE_URL=http://vllm:8000 PRISM_ADAPTER_DIR=/path/to/adapter \
  #     docker compose --profile serving up
  # (VLLM_BASE_URL must point at the service name so api/worker can reach it.)
  #
  # vllm:
  #   profiles: ["serving"]
  #   image: vllm/vllm-openai:v0.6.3
  #   restart: unless-stopped
  #   command: >
  #     --model meta-llama/Meta-Llama-3-8B-Instruct
  #     --enable-lora
  #     --lora-modules code-review=/adapter
  #     --served-model-name code-review-llama3-lora
  #     --host 0.0.0.0 --port 8000
  #   deploy:
  #     resources:
  #       reservations:
  #         devices:
  #           - driver: nvidia
  #             count: 1
  #             capabilities: [gpu]
  #   environment:
  #     HUGGING_FACE_HUB_TOKEN: ${HF_TOKEN:-}   # only if the base model is gated
  #   volumes:
  #     - ${PRISM_ADAPTER_DIR:-./adapter}:/adapter:ro
  #     - hf-cache:/root/.cache/huggingface
  #   healthcheck:
  #     test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5)"]
  #     interval: 30s
  #     timeout: 10s
  #     retries: 5
  #     start_period: 120s

volumes:
  redis-data:
  feedback-data:
  mlflow-data:
  # hf-cache:   # uncomment with the serving profile above