notifkit/profiling/docker-compose.yml

# Driven by src/runner.ts through the docker compose CLI; every ${VAR} below is
# set per tier by the runner. Nodes run the image the runner builds as
# notifkit-prof-node from profiling/Dockerfile.

name: notifkit-prof

x-node-env: &node-env
  NODE_ENV: production
  DATABASE_URL: postgres://notifkit:password@db:5432/notifkit
  REDIS_URL: redis://redis:6379
  ADMIN_API_KEY: perf-admin-key
  HOST: 0.0.0.0
  LOG_LEVEL: ${LOG_LEVEL:-info}
  WORKER_CONCURRENCY: ${WORKER_CONCURRENCY:-50}
  DELIVERY_CONCURRENCY: ${DELIVERY_CONCURRENCY:-200}
  DB_MAX_CONNECTIONS: ${DB_MAX_CONNECTIONS:-5}
  PIPELINE_FUSED: ${PIPELINE_FUSED:-false}
  RATE_LIMIT_PER_HOUR: ${RATE_LIMIT_PER_HOUR:-100}
  # Far above any backlog a run builds: MAXLEN trims unread entries, and a
  # trimmed notification is simply lost.
  QUEUE_MAX_LEN: ${QUEUE_MAX_LEN:-5000000}
  UNSUBSCRIBE_SECRET: profiling-unsubscribe-secret-0123456789
  PUBLIC_URL: http://localhost:3000
  PROVIDER_LATENCY_MS: ${PROVIDER_LATENCY_MS:-150-250}
  PROVIDER_FAILURE_RATE: ${PROVIDER_FAILURE_RATE:-0}
  # Share of delivered messages the runner later reports as opened through the
  # profiling transport's webhook (tests/sustained.test.ts).
  PROFILING_OPEN_RATE: ${PROFILING_OPEN_RATE:-0}
  # Worker services wait for the runner to open this gate, so ingest can be
  # measured on its own and the backlog it builds drained as a second phase.
  WORKER_GATE: "1"
  # Set by --cpu-prof: V8 profiles land in results/cpuprof on shutdown.
  NODE_OPTIONS: ${NODE_OPTIONS:-}

x-node: &node
  image: notifkit-prof-node:latest
  volumes:
    - ./results/cpuprof:/prof
  # Long enough for workers to finish in-flight messages and flush a profile.
  stop_grace_period: 60s
  depends_on:
    db:
      condition: service_healthy
    redis:
      condition: service_healthy
  healthcheck:
    test: ["CMD-SHELL", "wget -qO- http://127.0.0.1:3000/health >/dev/null || exit 1"]
    interval: 2s
    timeout: 2s
    retries: 60
    start_period: 5s

services:
  db:
    image: postgres:15-alpine
    environment:
      POSTGRES_DB: notifkit
      POSTGRES_USER: notifkit
      POSTGRES_PASSWORD: password
    # Production settings scaled to the tier's memory, rather than the image's
    # defaults (128MB shared_buffers, 1GB max_wal_size) that nobody deploys.
    command:
      - postgres
      - -c
      - shared_buffers=${PG_SHARED_BUFFERS:-256MB}
      - -c
      - effective_cache_size=${PG_EFFECTIVE_CACHE_SIZE:-768MB}
      - -c
      - maintenance_work_mem=${PG_MAINTENANCE_WORK_MEM:-64MB}
      - -c
      - work_mem=8MB
      - -c
      - max_connections=400
      - -c
      - wal_buffers=16MB
      - -c
      - min_wal_size=512MB
      - -c
      - max_wal_size=4GB
      - -c
      - checkpoint_completion_target=0.9
      - -c
      - random_page_cost=1.1
      - -c
      - effective_io_concurrency=200
      - -c
      - shared_preload_libraries=pg_stat_statements
      - -c
      - pg_stat_statements.track=all
      - -c
      - track_io_timing=on
    shm_size: 512mb
    ports:
      - "${PROF_DB_PORT:-35432}:5432"
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U notifkit -d notifkit"]
      interval: 1s
      timeout: 2s
      retries: 60
    deploy:
      resources:
        limits:
          cpus: "${DB_CPUS:-1.0}"
          memory: "${DB_MEMORY:-1G}"

  redis:
    image: redis:7-alpine
    # AOF with a once-a-second fsync is the durability production runs with;
    # the streams are the only copy of a queued notification.
    command:
      [
        "redis-server",
        "--appendonly",
        "yes",
        "--appendfsync",
        "everysec",
        "--save",
        "",
        "--maxmemory-policy",
        "noeviction",
      ]
    ports:
      - "${PROF_REDIS_PORT:-36379}:6379"
    healthcheck:
      test: ["CMD", "redis-cli", "ping"]
      interval: 1s
      timeout: 2s
      retries: 60
    deploy:
      resources:
        limits:
          cpus: "${REDIS_CPUS:-1.0}"
          memory: "${REDIS_MEMORY:-1G}"

  # ── Vertical: one monolith running every service ──────────────────────────
  server:
    <<: *node
    environment:
      <<: *node-env
      SERVICES: ${SERVER_SERVICES:-api,enricher,engine,delivery,scheduler,events}
    ports:
      - "${PROF_API_PORT:-35678}:3000"
    deploy:
      resources:
        limits:
          cpus: "${SERVER_CPUS:-1.0}"
          memory: "${SERVER_MEMORY:-1G}"

  # ── Horizontal: API ingress nodes ─────────────────────────────────────────
  api-1:
    <<: *node
    environment:
      <<: *node-env
      SERVICES: api
    ports:
      - "${PROF_API_PORT:-35678}:3000"
    deploy:
      resources:
        limits:
          cpus: "${API_CPUS:-1.0}"
          memory: "${API_MEMORY:-1G}"

  api-2:
    <<: *node
    environment:
      <<: *node-env
      SERVICES: api
    deploy:
      resources:
        limits:
          cpus: "${API_CPUS:-1.0}"
          memory: "${API_MEMORY:-1G}"

  # ── Horizontal: identical pipeline workers, scaled by count ───────────────
  worker-1: &worker
    <<: *node
    environment:
      <<: *node-env
      SERVICES: ${WORKER_SERVICES:-enricher,engine,delivery,scheduler,events}
    deploy:
      resources:
        limits:
          cpus: "${WORKER_CPUS:-1.0}"
          memory: "${WORKER_MEMORY:-1G}"

  worker-2: *worker
  worker-3: *worker
  worker-4: *worker

  # ── Load generator, inside the compose network ────────────────────────────
  loadgen:
    image: grafana/k6:1.0.0
    volumes:
      - ./k6:/scripts:ro
      - ./results/.k6:/out
    command: ["run", "--quiet", "--no-usage-report", "/scripts/notify.js"]
    deploy:
      resources:
        limits:
          cpus: "${LOADGEN_CPUS:-4.0}"
          memory: "2G"