doca-collectx-deployment

par nvidia

Utilisez cette compétence pour déployer et exploiter un collecteur de télémétrie DOCA basé sur CollectX (clx) sur un hôte ou un BlueField — connectant les providers / compteurs au collecteur,…

npx skills add https://github.com/nvidia/skills --skill doca-collectx-deployment

DOCA CollectX telemetry deployment

Where to start: This skill is the bundle's home for operating a CollectX (clx) based telemetry collector — the collection framework that gathers provider counters into a schema and ships them out through one or more exporters. It is a deployment / operation skill, parallel to doca-bare-metal-deployment and doca-container-deployment: it owns the runtime shape of a telemetry collector on the operator's host or BlueField, not the library APIs the operator's own program calls. If the user wants to stand up, wire, or debug a collector and its exporters, open TASKS.md and start at ## configure. If the question is what surfaces does the collector even have and where is the scope boundary, start at CAPABILITIES.md. If the user has not installed DOCA yet, route to doca-setup first.

The scope boundary (read this before anything else)

CollectX (clx) is NVIDIA's telemetry collection framework. It underpins the DOCA Telemetry Service (DTS) — and DTS as-deployed (the productized, NGC-shipped / kubelet-started service container) is out of scope for this bundle per AGENTS.md Non-goal #7. This skill therefore draws a hard line and the agent MUST state it up front:

  • In scope here: the CollectX collection mechanism as a class (providers / counters → schema → collector daemon → exporters), and the operator deploying / running / debugging a collector that they own, plus the operator's own usage of the two in-bundle telemetry libraries when those feed or consume the collector.
  • Routed to the DOCA telemetry libraries: the hardware-counter reader API is owned by doca-telemetry; the application-side publisher API (emit counters / events from a DOCA program) is owned by doca-telemetry-exporter. This skill does not re-document either API surface.
  • Routed to public docs (Non-goal #7): the productized DTS container — its packaged config schema, its built-in provider set, its kubelet manifest, its NGC image — is externally productized. Route every "operate the DTS service" question to the doca-public-knowledge-map externally-productized routing row and the public DTS guide it points at. The agent must NOT synthesize DTS config file names, provider knob names, or paths from memory.

The load-bearing first-touch failure this skill exists to prevent is collapsing these four surfaces into "DOCA telemetry": the clx collection mechanism, the doca-telemetry reader library, the doca-telemetry-exporter publisher library, and the productized DTS container are four different things with four different owners. The agent surfaces the decomposition BEFORE any config-level guidance.

Audience

This skill serves external operators standing up or running a CollectX-based telemetry collector on a host or BlueField they administer — people who already have:

  • a DOCA install on the side they are collecting from (host x86 or BlueField Arm), verified per doca-setup ## test,
  • a goal of getting counters off the box through a collector + exporter, not of writing the reader / publisher library code (that is the two libs/ skills above), and
  • access to the public DOCA Telemetry and DTS guides on docs.nvidia.com as the authoritative source for any concrete provider name, schema field, flag, or config path.

It is not for:

  • developers writing the hardware-counter reader API (route to doca-telemetry) or the publisher API (route to doca-telemetry-exporter),
  • operators deploying / configuring the productized DTS container as a turnkey service — that is externally productized (Non-goal #7); route to the public DTS guide,
  • fresh-no-install users — those belong on doca-setup ## no-install.

The skill teaches the agent the procedure and the scope boundary; it does not invent clx symbol names, provider names, schema field names, exporter flag names, or config paths from memory — those come from the live install and the public docs via doca-public-knowledge-map.

When to load this skill

Load this skill when the user is doing hands-on deployment or operation of a CollectX-based telemetry collector and the question is about the collector runtime shape, not a library API. Concretely:

  • Standing up a collector that gathers provider counters into a schema and ships them out — and deciding which export backend (Prometheus pull, Fluent Bit push, NetFlow, file / IPC) fits the downstream consumer.
  • Wiring a provider / counter family into the collector and confirming the device actually exposes it before the config commits (the gate-before-commit rule, shared with doca-telemetry-utils).
  • Turning on / shaping an exporter so the metrics actually leave the box, and confirming the downstream consumer receives them end-to-end (not just "the daemon is running").
  • Diagnosing a collector that starts but produces no schema rows, or ships nothing downstream, or whose exporter endpoint is silent — walking the layered ladder rather than guessing.
  • Recognising when the user is actually asking about the productized DTS container (route to public docs, Non-goal #7), the reader library (route to doca-telemetry), or the publisher library (route to doca-telemetry-exporter) instead of the collection mechanism this skill owns.

Do not load this skill for: the hardware-counter reader API (use doca-telemetry); the publisher API (use doca-telemetry-exporter); operating the productized DTS container (route via doca-public-knowledge-map Non-goal #7); installing DOCA or preparing the env (use doca-setup); or any hardware-state change (use doca-hardware-safety).

What this skill provides

This is a thin loader. The substantive material lives in two companion files:

  • CAPABILITIES.md — the collector deployment contract as a class: the four-surface decomposition (clx collection mechanism vs reader library vs publisher library vs productized DTS), the collection pipeline shape (providers / counters → schema → collector daemon → exporters), the export-backend surface (Prometheus pull, Fluent Bit push, NetFlow, file / IPC) at class level, the version overlay on doca-version, the error taxonomy (collector won't start → no provider rows → schema mismatch → exporter silent → downstream skew → transport), the observability surface, and the safety policy (gate provider support before commit; collector is read-only against the device; route any mutating step to doca-hardware-safety; do not invent clx names / paths).
  • TASKS.md — step-by-step workflows for the deployment verbs: configure, build (routing stub), modify, run, test, debug, plus a Deferred task verbs block that routes out-of-scope questions (the two libraries, the productized DTS container, env prep, hardware-state change) to their owners.

The skill assumes a host or BlueField where DOCA is already installed and healthy (per doca-setup ## test) and the operator can run the collector and reach its exporter sinks. It does not cover installing DOCA — that path goes through doca-setup — and it does not cover the reader / publisher library APIs or the productized DTS container.

Loading order

  1. Read this SKILL.md first to confirm the user's question is in scope (operating a CollectX-based collector and its exporters — NOT the reader / publisher library APIs, NOT the productized DTS container).
  2. For the four-surface decomposition, the collection pipeline shape, the export-backend class surface, the version overlay, the error taxonomy, the observability surface, and the safety policy, see CAPABILITIES.md.
  3. For step-by-step workflows — configure, build (routing stub), modify, run, test, debug, and the Deferred task verbs block — see TASKS.md.

Both companion files cross-link to each other, doca-version for the canonical version-handling rules, and doca-public-knowledge-map whenever the right answer is "read the live config / public docs" rather than collector-specific guidance.

Example questions this skill answers well

See references/details.md.

What this skill deliberately does not ship

See references/details.md.

Related skills

See references/details.md.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…