docker-compose.bench.yml

# Standalone bench stack — NOT an overlay on docker-compose.yml.
#
# Runs alongside the personal BrainDB stack with strict isolation at every
# layer: separate Docker project namespace, separate Postgres container,
# separate volume, separate ports, separate host data dir. The personal
# `braindb` database is never touched by anything here.
#
# Run with:
#   docker compose -f docker-compose.bench.yml --env-file .env.bench up
#
# Plain `docker compose up` runs the personal stack as normal — the bench
# stack cannot fire by accident.
#
# Safety properties (also enforced by code in benchmarks/beam/bench.py):
#   - Postgres database is named `braindb_bench` (never `braindb`)
#   - Postgres runs in its OWN container on host port 5434 (the user's
#     personal Postgres is already on 5433, so bench takes 5434 to avoid
#     a host-port collision)
#   - The bench BrainDB API is on host port 8001 (not 8000)
#   - The bench watcher mounts ./data_bench/ as /app/data, so it never
#     sees files in ./data/ (the personal watcher's directory)
#   - The bench runner asserts the active DATABASE_URL contains
#     `braindb_bench` before any destructive op; refuses otherwise

name: braindb_bench

services:
  postgres_bench:
    # pgvector + pg_trgm baked in; matches what alembic migrations expect.
    image: pgvector/pgvector:pg17
    container_name: braindb_bench_postgres
    restart: unless-stopped
    environment:
      POSTGRES_DB: braindb_bench
      POSTGRES_USER: braindb_bench
      # NOT a secret — bench-local only, never reachable from outside the
      # host. The hard-coded safety check in bench.py looks for the
      # literal "braindb_bench" substring in DATABASE_URL.
      POSTGRES_PASSWORD: bench_local_only
    ports:
      # Personal Postgres is already on 5433 in this environment; bench
      # takes 5434 to avoid a host-port collision. Verified before commit.
      - "5434:5432"
    volumes:
      - braindb_bench_pgdata:/var/lib/postgresql/data
    networks:
      - bench-network
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U braindb_bench -d braindb_bench"]
      interval: 5s
      timeout: 3s
      retries: 10

  api_bench:
    build: .
    container_name: braindb_bench_api
    restart: unless-stopped
    depends_on:
      postgres_bench:
        condition: service_healthy
    networks:
      - bench-network
    environment:
      # MUST contain "braindb_bench" OR "braindb_conv_" or the bench runner
      # refuses to start. The bench runner overrides this via the BENCH_DATABASE_URL
      # env var when it creates per-conversation databases (braindb_conv_001, ...).
      DATABASE_URL: ${BENCH_DATABASE_URL:-postgresql://braindb_bench:bench_local_only@postgres_bench:5432/braindb_bench}
      API_PORT: 8001
      HF_TOKEN: ${HF_TOKEN:-}
      # Bench defaults to local Qwen via the workstation tunnel — matches the
      # plan's Qwen-as-judge story. Override via .env.bench if needed.
      LLM_PROFILE: ${LLM_PROFILE_BENCH:-deepinfra}
      AGENT_MODEL: ${AGENT_MODEL:-}
      NVIDIA_NIM_API_KEY: ${NVIDIA_NIM_API_KEY:-}
      DEEPINFRA_API_KEY: ${DEEPINFRA_API_KEY:-}
      VLLM_API_KEY: ${VLLM_API_KEY:-EMPTY}
      # Verbose by default during bench — every tool call is logged so the
      # smoke tests can be audited.
      AGENT_VERBOSE: ${AGENT_VERBOSE_BENCH:-true}
      # Bench-specific tuning: wiki maintenance fires faster than prod's
      # 30-min freshness window so per-conversation warmup stays in minutes.
      WIKI_FRESHNESS_MINUTES: ${WIKI_FRESHNESS_MINUTES_BENCH:-1}
    extra_hosts:
      # Lets self-hosted profiles (vllm_workstation_qwen) reach a server
      # bound to the Docker host's loopback or the SSH tunnel.
      - "host.docker.internal:host-gateway"
    ports:
      - "8001:8001"
    volumes:
      # Bind the repo (for code reload) but OVERRIDE /app/data with the
      # bench-only directory. The personal watcher polls ./data/sources/;
      # the bench watcher polls ./data_bench/sources/. They cannot see
      # each other's files.
      - .:/app
      - ./data_bench:/app/data
    command: >
      sh -c "alembic upgrade head && uvicorn braindb.main:app --host 0.0.0.0 --port 8001"

  watcher_bench:
    build: .
    container_name: braindb_bench_watcher
    restart: unless-stopped
    depends_on:
      - api_bench
    networks:
      - bench-network
    environment:
      BRAINDB_API_URL: http://api_bench:8001
      # Faster poll than prod's 7s, so the smoke test feels responsive.
      INGEST_POLL_INTERVAL: ${INGEST_POLL_INTERVAL_BENCH:-3}
    volumes:
      - .:/app
      - ./data_bench:/app/data
    command: python -m braindb.ingest_watcher

  wiki_scheduler_bench:
    build: .
    container_name: braindb_bench_wiki_scheduler
    restart: unless-stopped
    depends_on:
      - api_bench
    networks:
      - bench-network
    environment:
      BRAINDB_API_URL: http://api_bench:8001
      # Wiki pipeline runs identical to prod — only the cadence changes.
      WIKI_ENABLED: 'true'
      WIKI_INTERVAL: ${WIKI_INTERVAL_BENCH:-5}
    volumes:
      - .:/app
    command: python -m braindb.wiki_scheduler

  # The bench_runner container is the orchestrator: it creates per-conversation
  # databases, restarts api_bench with each new DATABASE_URL, drops the
  # conversation .md into the watcher's dir, waits for warmup, asks the
  # probing questions, records answers. Uses the same image as api_bench (so
  # it has Python + BrainDB deps) plus the docker CLI installed at startup
  # and the docker socket mounted so it can recreate api_bench.
  #
  # Profile "runner" means it does NOT auto-start with `docker compose up`;
  # invoke it explicitly via:
  #   docker compose -f docker-compose.bench.yml run --rm bench_runner \
  #     python -m benchmarks.beam.bench --split 100K --limit 1
  # bench_runner — orchestrator for the BEAM bench. Creates per-conversation
  # Postgres databases, restarts api_bench with each new DATABASE_URL,
  # drops the .md into the watcher's dir, waits for warmup, asks the probing
  # questions. See benchmarks/beam/runner.Dockerfile.
  #
  # Profile "runner" means it does NOT auto-start with `docker compose up`.
  # Invoke it explicitly:
  #   docker compose -f docker-compose.bench.yml run --rm bench_runner \
  #     python -m benchmarks.beam.bench --split 100K --limit 1
  bench_runner:
    build:
      context: .
      dockerfile: benchmarks/beam/runner.Dockerfile
    container_name: braindb_bench_runner
    profiles: ["runner"]
    depends_on:
      postgres_bench:
        condition: service_healthy
    networks:
      - bench-network
    environment:
      # Bench-side runner reaches the api via Docker network DNS, no host port.
      BENCH_API_BASE: http://api_bench:8001
      # Per-conversation DB creation uses admin creds against postgres_bench's
      # built-in `postgres` maintenance DB.
      BENCH_ADMIN_DATABASE_URL: postgresql://braindb_bench:bench_local_only@postgres_bench:5432/postgres
      # bench.py mints per-conv URLs by appending `/braindb_conv_NNN` to this.
      BENCH_DB_BASE_URL: postgresql://braindb_bench:bench_local_only@postgres_bench:5432
      # Judge LLM (OpenAI-compatible endpoint). Empty default — Python config
      # in benchmarks/beam/config.py supplies the real default (deepinfra).
      # Set QWEN_BASE_URL_BENCH / QWEN_MODEL_BENCH in .env.bench to point at
      # a self-hosted vLLM endpoint (e.g. workstation Qwen on port 8010).
      QWEN_BASE_URL: ${QWEN_BASE_URL_BENCH:-}
      QWEN_MODEL: ${QWEN_MODEL_BENCH:-}
      QWEN_API_KEY: ${QWEN_API_KEY:-${VLLM_API_KEY:-}}
      HF_TOKEN: ${HF_TOKEN:-}
      # Docker CLI invocations need to know which compose file + project.
      COMPOSE_FILE: /app/docker-compose.bench.yml
      COMPOSE_PROJECT_NAME: braindb_bench
    extra_hosts:
      - "host.docker.internal:host-gateway"
    volumes:
      - .:/app
      - ./data_bench:/app/data
      # Docker socket lets the runner recreate api_bench between conversations.
      - /var/run/docker.sock:/var/run/docker.sock

volumes:
  braindb_bench_pgdata:
    name: braindb_bench_pgdata

networks:
  bench-network:
    name: braindb_bench_network