docker-compose.yml

services:
  gpu-monitor:
    build: .
    restart: unless-stopped
    runtime: nvidia          # requires NVIDIA Container Toolkit
    environment:
      # ── Notification channels (configure one or more) ──────────────────────
      - SLACK_WEBHOOK_URL=${SLACK_WEBHOOK_URL:-}
      - DISCORD_WEBHOOK_URL=${DISCORD_WEBHOOK_URL:-}
      - TELEGRAM_BOT_TOKEN=${TELEGRAM_BOT_TOKEN:-}
      - TELEGRAM_CHAT_ID=${TELEGRAM_CHAT_ID:-}
      - EMAIL_SMTP_HOST=${EMAIL_SMTP_HOST:-}
      - EMAIL_SMTP_PORT=${EMAIL_SMTP_PORT:-587}
      - EMAIL_USER=${EMAIL_USER:-}
      - EMAIL_PASS=${EMAIL_PASS:-}
      - EMAIL_TO=${EMAIL_TO:-}
      - TWILIO_ACCOUNT_SID=${TWILIO_ACCOUNT_SID:-}
      - TWILIO_AUTH_TOKEN=${TWILIO_AUTH_TOKEN:-}
      - TWILIO_FROM=${TWILIO_FROM:-}
      - TWILIO_TO=${TWILIO_TO:-}
      - WECOM_WEBHOOK_URL=${WECOM_WEBHOOK_URL:-}
      - FEISHU_WEBHOOK_URL=${FEISHU_WEBHOOK_URL:-}
      - DINGTALK_WEBHOOK_URL=${DINGTALK_WEBHOOK_URL:-}
      - BARK_URL=${BARK_URL:-}
      - NTFY_URL=${NTFY_URL:-}
      - NTFY_TOKEN=${NTFY_TOKEN:-}
      - GOTIFY_URL=${GOTIFY_URL:-}
      - GOTIFY_TOKEN=${GOTIFY_TOKEN:-}
      - PUSHOVER_TOKEN=${PUSHOVER_TOKEN:-}
      - PUSHOVER_USER=${PUSHOVER_USER:-}
      - ROCKETCHAT_WEBHOOK_URL=${ROCKETCHAT_WEBHOOK_URL:-}
      - MATTERMOST_WEBHOOK_URL=${MATTERMOST_WEBHOOK_URL:-}
      - TEAMS_WEBHOOK_URL=${TEAMS_WEBHOOK_URL:-}
      - GOOGLE_CHAT_WEBHOOK_URL=${GOOGLE_CHAT_WEBHOOK_URL:-}
      - ZULIP_SITE=${ZULIP_SITE:-}
      - ZULIP_EMAIL=${ZULIP_EMAIL:-}
      - ZULIP_API_KEY=${ZULIP_API_KEY:-}
      - OPENCLAW_WEBHOOK_URL=${OPENCLAW_WEBHOOK_URL:-}
      - OPENCLAW_WEBHOOK_SECRET=${OPENCLAW_WEBHOOK_SECRET:-}
      - PAGERDUTY_INTEGRATION_KEY=${PAGERDUTY_INTEGRATION_KEY:-}
      - APPRISE_URLS=${APPRISE_URLS:-}

      # ── Alerting thresholds ─────────────────────────────────────────────────
      - CHECK_INTERVAL=${CHECK_INTERVAL:-60}
      - IDLE_THRESHOLD=${IDLE_THRESHOLD:-10}
      - IDLE_MINUTES=${IDLE_MINUTES:-5}
      - ALERT_COOLDOWN=${ALERT_COOLDOWN:-30}
      - GPU_TEMP_WARN=${GPU_TEMP_WARN:-85}
      - GPU_TEMP_CRIT=${GPU_TEMP_CRIT:-92}
      - MEMLEAK_THRESHOLD=${MEMLEAK_THRESHOLD:-30}
      - MEMLEAK_MINUTES=${MEMLEAK_MINUTES:-10}
      - ALERT_WEBHOOK_URL=${ALERT_WEBHOOK_URL:-}

      # ── Metrics backends ────────────────────────────────────────────────────
      - WEB_PORT=${WEB_PORT:-8080}
      - INFLUXDB_URL=${INFLUXDB_URL:-}
      - INFLUXDB_TOKEN=${INFLUXDB_TOKEN:-}
      - INFLUXDB_BUCKET=${INFLUXDB_BUCKET:-gpu_metrics}
      - INFLUXDB_ORG=${INFLUXDB_ORG:-}
      - DATADOG_STATSD_HOST=${DATADOG_STATSD_HOST:-}
      - DATADOG_STATSD_PORT=${DATADOG_STATSD_PORT:-8125}
      - OTEL_EXPORTER_OTLP_ENDPOINT=${OTEL_EXPORTER_OTLP_ENDPOINT:-}
      - OTEL_SERVICE_NAME=${OTEL_SERVICE_NAME:-gpu-monitor}
      - OTEL_EXPORTER_OTLP_HEADERS=${OTEL_EXPORTER_OTLP_HEADERS:-}

      # ── GitHub Pages dashboard ─────────────────────────────────────────────
      - GITHUB_PAGES_TOKEN=${GITHUB_PAGES_TOKEN:-}
      - GITHUB_PAGES_REPO=${GITHUB_PAGES_REPO:-}

      - LOG_FILE=/var/log/gpu-monitor.log
    volumes:
      - gpu-monitor-logs:/var/log
    ports:
      - "${WEB_PORT:-8080}:${WEB_PORT:-8080}"
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:${WEB_PORT:-8080}/metrics', timeout=5)"]
      interval: 60s
      timeout: 10s
      retries: 3
      start_period: 30s

volumes:
  gpu-monitor-logs: