# ML4T 3rd Edition — Docker Compose Configuration
#
# QUICK START (pre-built images from Docker Hub):
#   docker compose pull ml4t         # Pull pre-built image (~12 GB amd64, ~3 GB arm64)
#   docker compose up ml4t           # Start Jupyter Lab at http://localhost:8888
#
# To build locally instead:
#   docker compose build ml4t        # Build from Dockerfile (slow, ~45 min)
#
# PLATFORM SUPPORT:
#   - Linux x86_64:     All services work, GPU optional
#   - Windows x86_64:   All services via Docker Desktop (WSL2), GPU via WSL2+nvidia-toolkit
#   - macOS Intel:      All services except GPU
#   - Apple Silicon:    ml4t + benchmark native arm64; ml4t-py312 amd64 only, runs under
#                       Rosetta emulation (see below)
#
# IMAGES ON DOCKER HUB (docker.io/ml4t/):
#   ml4t:latest           — All chapters, Python 3.14, PyTorch CUDA 12.8 (amd64+arm64)
#   ml4t-py312:latest     — Python 3.12: signatory, esig, gensim, tfcausalimpact (amd64 only)
#   ml4t-benchmark:latest — Storage benchmarks: DuckDB, HDF5, DB clients (amd64+arm64)
#
# WHICH IMAGE DO I NEED?
#   Most readers: ml4t only (covers Ch01-Ch27 + case studies)
#   Ch05 NB01/03/07, Ch09 NB06/12, Ch10 NB01-03, Ch12 NB10, Ch14 NB06, Ch15 NB06:
#       ml4t-py312 (amd64; on Apple Silicon read the committed .ipynb outputs, or run it
#       under Rosetta with DOCKER_DEFAULT_PLATFORM=linux/amd64)
#   Ch02 storage benchmarks: benchmark + database services
#   Ch12 GBM GPU benchmark: rapids (requires NVIDIA GPU)
#
# See docs/installation.md for detailed setup instructions

x-common: &common
  user: "${UID:-1000}:${GID:-1000}"
  volumes:
    - .:/app:rw
    # Read-write so the in-container download workflow (data/download_all.py)
    # can populate /data. ML4T_DATA_PATH=/data below points every downloader
    # here; a read-only mount makes each write fail with "Read-only file system".
    - ${ML4T_DATA_PATH:-./data}:/data:rw
    # Every notebook that pins a HuggingFace checkpoint loads it with
    # local_files_only=True, so a container run has to see the host's cache or the
    # load raises LocalEntryNotFoundError. Read-write, because the notebooks that do
    # not pin a checkpoint fetch it on first use (10_text_feature_engineering/09
    # loads FinBERT through a bare from_pretrained), and a read-only mount turns
    # that download into a failure. What a pinned load must not do is silently
    # substitute a different checkpoint, and `revision=` with local_files_only is
    # what prevents that, not the mount mode.
    #
    # The host directory has to exist before the first `docker compose` command.
    # Docker creates a missing bind source as root, and the container runs as the
    # invoking user, so an uncached download then fails on permissions despite the
    # :rw. `mkdir -p ~/.cache/huggingface` once is the whole fix; any host that has
    # run a HuggingFace library already has it.
    - ${HF_HOME:-${HOME}/.cache/huggingface}:/home/ml4t/.cache/huggingface:rw
  environment:
    - ML4T_DATA_PATH=/data
    - ML4T_PATH=/app
    - TEST=${TEST:-0}
    - NEO4J_URI=${NEO4J_URI:-bolt://neo4j:7687}
    - NEO4J_USER=${NEO4J_USER:-neo4j}
    - NEO4J_PASSWORD=${NEO4J_PASSWORD:-password}
    - NUMBA_CACHE_DIR=/tmp/numba_cache
    - MPLCONFIGDIR=/tmp/matplotlib
    - POLARS_FMT_MAX_ROWS=20
    - POLARS_FMT_STR_LEN=50
    - VIRTUAL_ENV=/opt/ml4t
    - UV_PROJECT_ENVIRONMENT=/opt/ml4t
    - PATH=/opt/ml4t/bin:/usr/local/bin:/usr/bin:/bin:/usr/local/games:/usr/games
  env_file:
    - path: .env
      required: false
  working_dir: /app
  stdin_open: true
  tty: true

services:
  # ============================================
  # ML4T MAIN (All Chapters)
  # Works on all platforms including Apple Silicon
  # ============================================
  ml4t:
    <<: *common
    image: ml4t/ml4t:latest
    build:
      context: .
      dockerfile: envs/ml4t/Dockerfile
    container_name: ml4t-review
    ports:
      - "127.0.0.1:8888:8888"
    command: ["jupyter", "lab", "--ip=0.0.0.0", "--port=8888", "--no-browser", "--allow-root"]

  # ============================================
  # ML4T GPU (NVIDIA only — Linux/Windows x86)
  # Same as ml4t but with GPU passthrough
  # ============================================
  ml4t-gpu:
    <<: *common
    image: ml4t/ml4t:latest
    build:
      context: .
      dockerfile: envs/ml4t/Dockerfile
    container_name: ml4t-review-gpu
    ports:
      - "127.0.0.1:8889:8888"
    command: ["jupyter", "lab", "--ip=0.0.0.0", "--port=8888", "--no-browser", "--allow-root"]
    profiles: ["gpu"]
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # ============================================
  # BENCHMARK (ARM64 Compatible)
  # Storage benchmarks: Parquet, DuckDB, HDF5, database clients
  # Excludes ArcticDB (no ARM64 wheels)
  # Works on Apple Silicon
  # ============================================
  benchmark:
    <<: *common
    image: ml4t/ml4t-benchmark:latest
    build:
      context: .
      dockerfile: envs/benchmark/Dockerfile
    container_name: ml4t-benchmark
    ports:
      - "127.0.0.1:8887:8888"
    profiles: ["benchmark"]
    # Re-declares the x-common volumes because a service-level `volumes:` key
    # replaces the anchor's rather than extending it.
    volumes:
      - .:/app:rw
      - ${ML4T_DATA_PATH:-./data}:/data:rw
      # kdb+/PyKX licence + q binary, mounted read-only from the host at RUNTIME.
      # Never baked into the image: the licence is personal and must not ship in
      # a published image. If you have no kdb+ licence these resolve to empty
      # directories, PyKX finds no licence, and 21_storage_benchmark_database
      # skips kdb+ and says so. Get a free personal edition from
      # https://kx.com/kdb-personal-edition-download/ and unpack it to ~/.kx.
      - ${KX_HOME:-${HOME}/.kx}:/home/ml4t/.kx:ro
      - ${PYKX_HOME:-${HOME}/.pykx}:/home/ml4t/.pykx:ro
    environment:
      - ML4T_DATA_PATH=/data
      - ML4T_PATH=/app
      - TEST=${TEST:-0}
      # q resolves its licence from QLIC, not from the file's presence on disk;
      # without this q starts and immediately fails "no license loaded".
      - QLIC=/home/ml4t/.kx
      - QHOME=/home/ml4t/.kx/q
      # Database connections (when running with database services)
      - CLICKHOUSE_HOST=ml4t-clickhouse
      - CLICKHOUSE_PORT=8123
      - QUESTDB_HOST=ml4t-questdb
      - QUESTDB_HTTP_PORT=9000
      - TIMESCALE_HOST=ml4t-timescaledb
      - TIMESCALE_PORT=5432
      - TIMESCALE_PASSWORD=benchmark
      - INFLUXDB_HOST=ml4t-influxdb
      - INFLUXDB_PORT=8086
      - INFLUXDB_ORG=ml4t
      - INFLUXDB_TOKEN=benchmark-token
      - INFLUXDB_BUCKET=market_data
      - POSTGRES_HOST=ml4t-postgres
      - POSTGRES_PORT=5432
      - POSTGRES_USER=postgres
      - POSTGRES_PASSWORD=benchmark
      - POSTGRES_DB=ml4t

  # ============================================
  # BENCHMARK FULL (x86 Only)
  # All benchmarks including ArcticDB
  # WARNING: Does not work on Apple Silicon
  # ============================================
  benchmark-full:
    <<: *common
    build:
      context: .
      dockerfile: envs/benchmark/Dockerfile.full
      # Force x86 build (required for ArcticDB)
      platforms:
        - linux/amd64
    container_name: ml4t-benchmark-full
    ports:
      - "127.0.0.1:8887:8888"
    profiles: ["benchmark-full"]
    environment:
      - ML4T_DATA_PATH=/data
      - ML4T_PATH=/app
      - TEST=${TEST:-0}
      # Database connections
      - CLICKHOUSE_HOST=ml4t-clickhouse
      - CLICKHOUSE_PORT=8123
      - QUESTDB_HOST=ml4t-questdb
      - QUESTDB_HTTP_PORT=9000
      - TIMESCALE_HOST=ml4t-timescaledb
      - TIMESCALE_PORT=5432
      - TIMESCALE_PASSWORD=benchmark
      - INFLUXDB_HOST=ml4t-influxdb
      - INFLUXDB_PORT=8086
      - INFLUXDB_ORG=ml4t
      - INFLUXDB_TOKEN=benchmark-token
      - INFLUXDB_BUCKET=market_data
      - POSTGRES_HOST=ml4t-postgres
      - POSTGRES_PORT=5432
      - POSTGRES_USER=postgres
      - POSTGRES_PASSWORD=benchmark
      - POSTGRES_DB=ml4t

  # ============================================
  # RAPIDS GBM BENCHMARK (NVIDIA GPU Required)
  # Ch12 GBM library benchmark: RAPIDS cuML XGBoost, LightGBM CUDA,
  # CatBoost GPU, plus CPU baselines. Single-purpose image.
  # Start with: docker compose --profile rapids run --rm rapids python 12_gradient_boosting/02_gbm_comparison.py
  # ============================================
  rapids:
    <<: *common
    build:
      context: .
      dockerfile: envs/rapids/Dockerfile
      platforms:
        - linux/amd64
    container_name: ml4t-rapids
    profiles: ["rapids"]
    user: root
    entrypoint: []
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # ============================================
  # PY312 (Python 3.12, x86 Only)
  # For packages without Python 3.14 support: gensim, signatory, esig,
  # tfcausalimpact, plus notebooks hitting the Python 3.14 torch CUDA import bug.
  # Covers: Ch05 NB01/03/07, Ch09 NB06/12, Ch10 NB01-03, Ch12 NB10, Ch14 NB06,
  # Ch15 NB06. Ch21 deep hedging is NOT here: pfhedge is a main dependency and
  # 0.22.0 works on Python 3.14, so that notebook runs in the default image.
  # Start with: docker compose --profile py312 run --rm py312 python ...
  # This service reserves no GPU, so it starts on macOS and on any machine
  # without NVIDIA. Use the py312-gpu service below where a GPU is available.
  # Apple Silicon: there is no native arm64 image (signatory/esig/gensim have no
  # arm64 wheels). Either read the committed .ipynb outputs for the notebooks
  # above, or run the amd64 image under Rosetta emulation (slow, no GPU):
  #   DOCKER_DEFAULT_PLATFORM=linux/amd64 docker compose --profile py312 run --rm py312 python ...
  # ============================================
  py312:
    <<: *common
    image: ml4t/ml4t-py312:latest
    build:
      context: .
      dockerfile: envs/py312/Dockerfile
      platforms:
        - linux/amd64
    container_name: ml4t-py312
    profiles: ["py312"]

  # Same image with an NVIDIA GPU attached, for the six GPU-tagged py312 notebooks
  # (Ch05 NB01/03/07, Ch10 NB03, Ch12 NB10, Ch14 NB06), whether they train or run
  # inference. Linux and Windows WSL2 only, exactly like ml4t-gpu.
  # Start with: docker compose --profile py312-gpu run --rm py312-gpu python ...
  py312-gpu:
    <<: *common
    image: ml4t/ml4t-py312:latest
    build:
      context: .
      dockerfile: envs/py312/Dockerfile
      platforms:
        - linux/amd64
    container_name: ml4t-py312-gpu
    profiles: ["py312-gpu"]
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

  # ============================================
  # KNOWLEDGE GRAPH (Neo4j Community Edition)
  # Start with: docker compose --profile kg up -d neo4j
  # ============================================
  neo4j:
    image: neo4j:2026.02.2
    container_name: ml4t-neo4j
    profiles: ["kg"]
    ports:
      - "127.0.0.1:7474:7474"
      - "127.0.0.1:7687:7687"
    environment:
      - NEO4J_AUTH=${NEO4J_AUTH:-neo4j/password}
    volumes:
      - neo4j_data:/data
      - neo4j_logs:/logs
    healthcheck:
      test: ["CMD-SHELL", "cypher-shell -u neo4j -p password 'RETURN 1;'"]
      interval: 10s
      timeout: 10s
      retries: 10

  # ============================================
  # Database Services (for benchmarks)
  # Start with: docker compose --profile databases up -d
  # ============================================
  timescaledb:
    image: timescale/timescaledb:latest-pg16
    container_name: ml4t-timescaledb
    environment:
      - POSTGRES_PASSWORD=benchmark
      - POSTGRES_DB=ml4t
      - POSTGRES_USER=postgres
    ports:
      - "127.0.0.1:5437:5432"
    profiles: ["benchmark", "benchmark-full", "databases"]
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U postgres"]
      interval: 10s
      timeout: 5s
      retries: 5

  postgres:
    image: postgres:16
    container_name: ml4t-postgres
    environment:
      - POSTGRES_PASSWORD=benchmark
      - POSTGRES_DB=ml4t
      - POSTGRES_USER=postgres
    ports:
      - "127.0.0.1:5436:5432"
    profiles: ["benchmark", "benchmark-full", "databases"]
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U postgres"]
      interval: 10s
      timeout: 5s
      retries: 5

  clickhouse:
    image: clickhouse/clickhouse-server:latest
    container_name: ml4t-clickhouse
    environment:
      - CLICKHOUSE_DEFAULT_ACCESS_MANAGEMENT=1
      - CLICKHOUSE_USER=default
      - CLICKHOUSE_PASSWORD=
    ports:
      - "127.0.0.1:8123:8123"
    profiles: ["benchmark", "benchmark-full", "databases"]
    deploy:
      resources:
        limits:
          memory: 4G
    healthcheck:
      test: ["CMD", "wget", "--spider", "-q", "http://localhost:8123/ping"]
      interval: 10s
      timeout: 5s
      retries: 5

  questdb:
    image: questdb/questdb:latest
    container_name: ml4t-questdb
    ports:
      - "127.0.0.1:9000:9000"
      - "127.0.0.1:9009:9009"
    profiles: ["benchmark", "benchmark-full", "databases"]
    deploy:
      resources:
        limits:
          memory: 2G
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:9000"]
      interval: 10s
      timeout: 5s
      retries: 5

  influxdb:
    image: influxdb:2.7
    container_name: ml4t-influxdb
    environment:
      - DOCKER_INFLUXDB_INIT_MODE=setup
      - DOCKER_INFLUXDB_INIT_USERNAME=admin
      - DOCKER_INFLUXDB_INIT_PASSWORD=benchmark123
      - DOCKER_INFLUXDB_INIT_ORG=ml4t
      - DOCKER_INFLUXDB_INIT_BUCKET=market_data
      - DOCKER_INFLUXDB_INIT_ADMIN_TOKEN=benchmark-token
    ports:
      - "127.0.0.1:8086:8086"
    profiles: ["benchmark", "benchmark-full", "databases"]
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8086/health"]
      interval: 10s
      timeout: 5s
      retries: 5

networks:
  default:
    name: ml4t-review-network

volumes:
  neo4j_data:
  neo4j_logs: