docker-compose.yml

services:
  db:
    image: postgres:18-bookworm
    restart: unless-stopped
    shm_size: '2GB'                    # Required for REFRESH MATERIALIZED VIEW with
                                      # maintenance_work_mem=2GB. Default Docker
                                      # shm (64MB) causes "No space left on device".
    environment:
      POSTGRES_DB: gnaf
      POSTGRES_USER: postgres
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?POSTGRES_PASSWORD must be set}
      # C locale is REQUIRED for text_pattern_ops index performance.
      # The Tier 0b and Tier 1 indexes use text_pattern_ops for prefix
      # matching. Without C locale, they fall back to ICU/UTF-8 which
      # adds 10-20% overhead per index comparison.
      POSTGRES_INITDB_ARGS: "--locale=C --encoding=UTF8"
    # No host port mapping — DB is only reachable via Docker internal network.
    # api connects at postgresql://postgres:${POSTGRES_PASSWORD}@db:5432/gnaf.
    volumes:
      # Postgres 18+ requires the mount at /var/lib/postgresql (not /data).
      # The image creates a major-version-specific subdirectory automatically.
      - pgdata:/var/lib/postgresql
      - ./postgresql.conf:/etc/postgresql/postgresql.conf
      # Mount individual SQL files (NOT 005_prewarm.sql — that runs after the
      # loader via scripts/load.ts, since the MV must have data first)
      - ./sql/001_extensions.sql:/docker-entrypoint-initdb.d/001_extensions.sql
      - ./sql/003b_abbrev_map.sql:/docker-entrypoint-initdb.d/003b_abbrev_map.sql
      - ./sql/003c_expand_fn.sql:/docker-entrypoint-initdb.d/003c_expand_fn.sql
      - ./sql/005_staging.sql:/docker-entrypoint-initdb.d/005_staging.sql
      - ./sql/006_staging_indexes.sql:/docker-entrypoint-initdb.d/006_staging_indexes.sql
      - ./sql/007_mv.sql:/docker-entrypoint-initdb.d/007_mv.sql
      - ./sql/008_api_keys.sql:/docker-entrypoint-initdb.d/008_api_keys.sql
      - ./sql/009_domain_verify.sql:/docker-entrypoint-initdb.d/009_domain_verify.sql
      - ./sql/010_last_verified.sql:/docker-entrypoint-initdb.d/010_last_verified.sql
      - ./sql/011_gnaf_roles.sql:/docker-entrypoint-initdb.d/011_gnaf_roles.sql
      - ./sql/012_api_key_expiry.sql:/docker-entrypoint-initdb.d/012_api_key_expiry.sql
    command: postgres -c config_file=/etc/postgresql/postgresql.conf
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U postgres -d gnaf"]
      interval: 5s
      timeout: 5s
      retries: 10
      start_period: 30s
    deploy:
      resources:
        limits:
          memory: 20G

  api:
    build: .
    restart: unless-stopped
    # CD pushes to GHCR; image tag is set by the CD workflow.
    # Local `docker compose up` builds from source (build: . takes precedence).
    image: ghcr.io/${GH_USER:-lst97}/gnaf-autocomplete:${IMAGE_TAG:-latest}
    ports:
      - "8000:8000"
    env_file: .env
    environment:
      # Docker-network-specific — overrides whatever DATABASE_URL the .env file sets,
      # because within the compose network Postgres is at `db:5432`, not localhost.
      DATABASE_URL: postgresql://postgres:${POSTGRES_PASSWORD}@db:5432/gnaf
      # NODE_ENV=production enables 120/min rate limiting.
      # Set via .env for production; omitted locally to keep it disabled.
      # NODE_ENV: ${NODE_ENV:-development}
      PUBLIC_URL: ${PUBLIC_URL:-}     # Same-origin domain mismatch bypass
    volumes:
      # G-NAF data directory — only needed by the loader, not by the API runtime.
      # The CD workflow auto-downloads data here if missing. Mounted rw so
      # the loader can create and read PSV files. Defaults to ../gnaf-data
      # (parent of the project root) for both local dev and production.
      # Override via GNAF_DATA_ROOT env var.
      - ${GNAF_DATA_ROOT:-../gnaf-data}:/opt/gnaf-data:rw
    depends_on:
      db:
        condition: service_healthy
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/healthz"]
      interval: 10s
      timeout: 5s
      retries: 3
      start_period: 10s
    deploy:
      resources:
        limits:
          # 8GB required for 9-worker parallel COPY during `bun run scripts/load.ts`.
          # Runtime API typically uses <500MB; the extra is for the loader which
          # runs 9 concurrent COPY FROM STDIN streams for the largest states.
          memory: 8G

  tunnel:
    image: cloudflare/cloudflared:latest
    command: tunnel --no-autoupdate run
    environment:
      - TUNNEL_TOKEN=${CF_TUNNEL_TOKEN}
    depends_on:
      - api
    restart: unless-stopped
    profiles:
      - production

volumes:
  pgdata: