{"id":51087124,"url":"https://github.com/cadaverine/dgx-spark-observability","last_synced_at":"2026-06-23T22:33:23.585Z","repository":{"id":356823534,"uuid":"1233954786","full_name":"cadaverine/dgx-spark-observability","owner":"cadaverine","description":"Layer-1 host observability template for DGX Spark / Blackwell GB10 / sm_121a / aarch64 — LGTM stack with documented Blackwell+ARM quirks. MIT.","archived":false,"fork":false,"pushed_at":"2026-05-09T22:18:30.000Z","size":582,"stargazers_count":0,"open_issues_count":2,"forks_count":0,"subscribers_count":0,"default_branch":"master","last_synced_at":"2026-05-10T00:29:04.938Z","etag":null,"topics":["alertmanager","arm64","blackwell","dgx-spark","docker-compose","gb10","grafana","lgtm-stack","loki","monitoring","ntfy","nvidia","observability","prometheus"],"latest_commit_sha":null,"homepage":null,"language":"Shell","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/cadaverine.png","metadata":{"files":{"readme":"README.md","changelog":"CHANGELOG.md","contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":"CODE_OF_CONDUCT.md","threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":"SECURITY.md","support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-05-09T15:03:13.000Z","updated_at":"2026-05-09T22:18:27.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/cadaverine/dgx-spark-observability","commit_stats":null,"previous_names":["cadaverine/dgx-spark-observability"],"tags_count":1,"template":false,"template_full_name":null,"purl":"pkg:github/cadaverine/dgx-spark-observability","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/cadaverine%2Fdgx-spark-observability","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/cadaverine%2Fdgx-spark-observability/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/cadaverine%2Fdgx-spark-observability/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/cadaverine%2Fdgx-spark-observability/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/cadaverine","download_url":"https://codeload.github.com/cadaverine/dgx-spark-observability/tar.gz/refs/heads/master","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/cadaverine%2Fdgx-spark-observability/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34709804,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-23T02:00:07.161Z","response_time":65,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["alertmanager","arm64","blackwell","dgx-spark","docker-compose","gb10","grafana","lgtm-stack","loki","monitoring","ntfy","nvidia","observability","prometheus"],"created_at":"2026-06-23T22:33:21.019Z","updated_at":"2026-06-23T22:33:23.580Z","avatar_url":"https://github.com/cadaverine.png","language":"Shell","funding_links":[],"categories":[],"sub_categories":[],"readme":"# dgx-spark-observability\n\nLayer-1 host observability for DGX Spark / Blackwell GB10 / sm_121a / aarch64.\n\n## What you get\n\nA self-contained, 11-container LGTM-style stack:\n\n- **Prometheus** + **Grafana** + **Loki** + **Promtail** + **Alertmanager** + **ntfy** (self-hosted push).\n- Exporters: **node-exporter**, **smartctl-exporter** (arm64-rebuilt), **cAdvisor**, **dcgm-exporter**, plus an **nvidia-textfile** sidecar that fills DCGM gaps on Blackwell unified memory.\n- 6-month metric retention, 90-day system-log retention.\n- 13 hardware alerts (RAM/disk/GPU temp/throttle/OOM/NVMe SMART/etc.) with runbook annotations.\n- Overnight push mute (23:00–08:00 in `${TZ}`).\n- Custom **38-panel** at-a-glance host dashboard (`Spark Host Overview`) plus three pre-wired Grafana community dashboards.\n\n## Why this exists\n\nGeneric LGTM templates assume x86_64 + discrete-VRAM NVIDIA GPUs. DGX Spark is **aarch64 + Blackwell GB10 with unified memory**, and that combination breaks several common assumptions:\n\n- `smartctl_exporter` ships **amd64-only** manifests upstream.\n- `dcgm-exporter` 4.x doesn't expose `FB_USED/FB_FREE` on unified memory and doesn't load the profiling module on sm_121a.\n- `cadvisor` v0.49.x uses a Docker client API too old for Docker 25+.\n- Alertmanager rejects overnight mute windows directly.\n- A handful of `nvidia-smi` query fields behave differently on unified memory and on driver 580+.\n\nThis template documents and works around all of these. See [`docs/DGX-SPARK-QUIRKS.md`](docs/DGX-SPARK-QUIRKS.md) for the full list with symptom / cause / fix / upstream status.\n\n## Hardware compatibility\n\n| | |\n|---|---|\n| **Tested** | DGX Spark (GB10 / sm_121a / aarch64), Ubuntu 24.04, Docker 25+ |\n| **Likely compatible** | NVIDIA Jetson Orin / Thor (aarch64 + NVIDIA), Grace+Hopper (aarch64 + Hopper), any ARM64 Linux + NVIDIA GPU |\n| **Untested but should work** | x86_64 NVIDIA hosts. You can swap `namerci/smartctl-exporter` for the upstream `prometheus-community/smartctl_exporter` and pick any DCGM tag you prefer. |\n\n## Quick start\n\n```bash\ngit clone https://github.com/cadaverine/dgx-spark-observability.git\ncd dgx-spark-observability\n\ncp .env.example .env\n# Edit .env: set GRAFANA_ADMIN_PASSWORD, NTFY_TOPIC (must be unguessable),\n#            BIND_ADDRESS, TZ.\n\ndocker network create observability-net\ndocker compose up -d\n# Wait ~60s for everything to be ready.\n\ncurl -sf http://127.0.0.1:3001/api/health   # Grafana check\n\n# Optional: download three Grafana community dashboards\n./scripts/fetch-community-dashboards.sh\ndocker compose restart grafana\n```\n\n## What it looks like\n\n\u003c!-- TODO: drop a real screenshot at docs/img/dashboard.png. The link above expects that path. --\u003e\n![Spark Host Overview dashboard](docs/img/dashboard.png)\n\n*38-panel dashboard covering system / GPU / thermal / storage / network / containers + logs in one view. See [`grafana/dashboards/custom/spark-host-overview.json`](grafana/dashboards/custom/spark-host-overview.json).*\n\n## Endpoints\n\nAll UIs bind to `${BIND_ADDRESS}` (default `127.0.0.1`). For remote access either set `BIND_ADDRESS` to a Tailscale IP, or SSH-forward:\n\n| Service | Port | URL |\n|---|---|---|\n| Grafana | `3001` | `http://${BIND_ADDRESS}:3001` (admin / `${GRAFANA_ADMIN_PASSWORD}`) |\n| Prometheus | `9090` | `http://${BIND_ADDRESS}:9090` |\n| Alertmanager | `9093` | `http://${BIND_ADDRESS}:9093` |\n| ntfy | `8081` | `http://${BIND_ADDRESS}:8081` (subscribe via the ntfy mobile app to topic `${NTFY_TOPIC}`) |\n\n## Architecture\n\nSee [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) for the full design — network model, metric sources, label schema, log pipeline, dashboard structure, alerting design, retention choices.\n\n## Quirks\n\nSee [`docs/DGX-SPARK-QUIRKS.md`](docs/DGX-SPARK-QUIRKS.md) for the Blackwell / aarch64 specific gotchas this template handles.\n\n## Customization\n\nThe template ships with two static labels — `host=spark`, `hardware=gb10` — applied in three places:\n\n- `prometheus/prometheus.yml` (`external_labels` + per-job `instance`)\n- `promtail/promtail-config.yml` (per-job `labels`)\n- Some alert annotations in `prometheus/rules/host.yml`\n\nIf you're running on something other than a DGX Spark, search-and-replace `spark` and `gb10` to whatever fits your inventory. The Grafana dashboard's `$host` template variable picks them up automatically.\n\n## What is NOT included\n\nThis template is intentionally **host-level only**. If you also want application-level monitoring, fork or extend:\n\n- vLLM `/metrics` scrape (TTFT, KV cache, queue, throughput).\n- LiteLLM / per-key cost tracking.\n- Container `stdout`/`stderr` shipping to Loki.\n- Distributed tracing (Tempo, etc.).\n- Multi-host federation, HA replicas, SSO.\n\nThe label schema and Prometheus / Loki are deliberately not pre-bound to any application so that adding these on top does not require rewriting Layer 1.\n\n## Contributing\n\nPRs welcome — particularly for additional hardware compatibility (Jetson, Grace+Hopper) or further quirk documentation.\n\nLicense: MIT.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fcadaverine%2Fdgx-spark-observability","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fcadaverine%2Fdgx-spark-observability","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fcadaverine%2Fdgx-spark-observability/lists"}