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
- 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).
- 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.
- 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.
1---2license: Apache-2.03name: doca-collectx-deployment4description: >5 Use this skill to deploy and operate a CollectX (clx) based6 DOCA telemetry collector on a host or BlueField — wiring7 providers / counters into the collector, running the8 collection daemon, and shaping its exporters (Prometheus9 pull, Fluent Bit push, NetFlow, file / IPC) so the metrics10 actually leave the box. Trigger even when11 the user never says CollectX or clx — implicit phrasings:12 {collector emits nothing downstream}, {add a provider to the13 clx collector}, {turn on the Prometheus endpoint}, {ship14 counters to Fluent Bit from the DPU}, {daemon starts but no15 schema rows appear}. This skill owns the CollectX collection16 mechanism plus the operator's own doca-telemetry /17 doca-telemetry-exporter usage; it ROUTES the18 productized DOCA Telemetry Service (DTS) to public docs19 (AGENTS.md Non-goal #7), the reader API to doca-telemetry, and20 the publisher API to doca-telemetry-exporter. Refuse to invent21 clx symbols, provider names, schema fields, flags, or config22 paths — describe the class and route to the live source.23metadata:24 source: https://skillmd.com/skills/nvidia/doca-collectx-deployment25 kind: library26compatibility: >27 No DOCA install required to read this skill (it is a28 deployment / operation overlay over the DOCA telemetry29 libraries and the CollectX collection mechanism). The30 hands-on steps DO require a live DOCA install at31 /opt/mellanox/doca on a host or BlueField, an operator32 account that can run the collector and reach its exporter33 sinks, and the public DOCA Telemetry / DTS guides on34 docs.nvidia.com for any concrete provider name, schema field,35 flag, or config path (this skill never invents those).36---3738# DOCA CollectX telemetry deployment3940**Where to start:** This skill is the bundle's home for41*operating a CollectX (clx) based telemetry collector* — the42collection framework that gathers provider counters into a43schema and ships them out through one or more exporters. It is a44deployment / operation skill, parallel to45[`doca-bare-metal-deployment`](../doca-bare-metal-deployment/SKILL.md)46and [`doca-container-deployment`](../doca-container-deployment/SKILL.md):47it owns the *runtime shape* of a telemetry collector on the48operator's host or BlueField, not the library APIs the operator's49own program calls. If the user wants to stand up, wire, or debug50a collector and its exporters, open [`TASKS.md`](TASKS.md) and51start at [`## configure`](TASKS.md#configure). If the question is52*what surfaces does the collector even have and where is the53scope boundary*, start at [`CAPABILITIES.md`](CAPABILITIES.md).54If the user has not installed DOCA yet, route to55[`doca-setup`](../doca-setup/SKILL.md) first.5657## The scope boundary (read this before anything else)5859CollectX (clx) is NVIDIA's telemetry **collection** framework. It60underpins the **DOCA Telemetry Service (DTS)** — and DTS61*as-deployed* (the productized, NGC-shipped / kubelet-started62service container) is **out of scope** for this bundle per63[`AGENTS.md` Non-goal #7](../../AGENTS.md#non-goals-questions-the-agent-should-recognize-and-refuse-politely).64This skill therefore draws a hard line and the agent MUST state65it up front:6667- **In scope here:** the CollectX *collection mechanism* as a68 class (providers / counters → schema → collector daemon →69 exporters), and the operator deploying / running / debugging a70 collector that they own, plus the operator's own usage of the71 two in-bundle telemetry **libraries** when those feed or72 consume the collector.73- **Routed to the DOCA telemetry libraries:** the74 hardware-counter **reader** API is owned by75 [`doca-telemetry`](../libs/doca-telemetry/SKILL.md); the76 application-side **publisher** API (emit counters / events from77 a DOCA program) is owned by78 [`doca-telemetry-exporter`](../libs/doca-telemetry-exporter/SKILL.md).79 This skill does not re-document either API surface.80- **Routed to public docs (Non-goal #7):** the productized81 **DTS container** — its packaged config schema, its built-in82 provider set, its kubelet manifest, its NGC image — is83 externally productized. Route every "operate the DTS service"84 question to the85 [`doca-public-knowledge-map` externally-productized routing row](../doca-public-knowledge-map/SKILL.md#externally-productized-doca-software--not-in-this-bundle-but-here-is-where-to-route)86 and the public DTS guide it points at. The agent must NOT87 synthesize DTS config file names, provider knob names, or88 paths from memory.8990The load-bearing first-touch failure this skill exists to91prevent is **collapsing these four surfaces into "DOCA92telemetry"**: the clx collection mechanism, the `doca-telemetry`93reader library, the `doca-telemetry-exporter` publisher library,94and the productized DTS container are four different things with95four different owners. The agent surfaces the decomposition96BEFORE any config-level guidance.9798## Audience99100This skill serves **external operators standing up or running a101CollectX-based telemetry collector** on a host or BlueField they102administer — people who already have:103104- a DOCA install on the side they are collecting from (host x86105 or BlueField Arm), verified per106 [`doca-setup ## test`](../doca-setup/TASKS.md#test),107- a goal of *getting counters off the box* through a collector +108 exporter, not of writing the reader / publisher library code109 (that is the two `libs/` skills above), and110- access to the public DOCA Telemetry and DTS guides on111 `docs.nvidia.com` as the authoritative source for any concrete112 provider name, schema field, flag, or config path.113114It is **not** for:115116- developers writing the hardware-counter reader API (route to117 [`doca-telemetry`](../libs/doca-telemetry/SKILL.md)) or the118 publisher API (route to119 [`doca-telemetry-exporter`](../libs/doca-telemetry-exporter/SKILL.md)),120- operators deploying / configuring the **productized DTS121 container** as a turnkey service — that is externally122 productized (Non-goal #7); route to the public DTS guide,123- fresh-no-install users — those belong on124 [`doca-setup ## no-install`](../doca-setup/TASKS.md#no-install).125126The skill teaches the agent the *procedure and the scope127boundary*; it does not invent clx symbol names, provider names,128schema field names, exporter flag names, or config paths from129memory — those come from the live install and the public docs via130[`doca-public-knowledge-map`](../doca-public-knowledge-map/SKILL.md).131132## When to load this skill133134Load this skill when the user is doing hands-on **deployment or135operation of a CollectX-based telemetry collector** and the136question is about the collector runtime shape, not a library API.137Concretely:138139- Standing up a collector that gathers provider counters into a140 schema and ships them out — and deciding which export backend141 (Prometheus pull, Fluent Bit push, NetFlow, file / IPC) fits142 the downstream consumer.143- Wiring a provider / counter family into the collector and144 confirming the device actually exposes it before the config145 commits (the gate-before-commit rule, shared with146 [`doca-telemetry-utils`](../tools/doca-telemetry-utils/SKILL.md)).147- Turning on / shaping an exporter so the metrics actually leave148 the box, and confirming the downstream consumer receives them149 end-to-end (not just "the daemon is running").150- Diagnosing a collector that starts but produces no schema rows,151 or ships nothing downstream, or whose exporter endpoint is152 silent — walking the layered ladder rather than guessing.153- Recognising when the user is actually asking about the154 productized DTS container (route to public docs, Non-goal #7),155 the reader library (route to `doca-telemetry`), or the156 publisher library (route to `doca-telemetry-exporter`) instead157 of the collection mechanism this skill owns.158159Do **not** load this skill for: the hardware-counter reader API160(use [`doca-telemetry`](../libs/doca-telemetry/SKILL.md)); the161publisher API (use162[`doca-telemetry-exporter`](../libs/doca-telemetry-exporter/SKILL.md));163operating the productized DTS container (route via164[`doca-public-knowledge-map`](../doca-public-knowledge-map/SKILL.md)165Non-goal #7); installing DOCA or preparing the env (use166[`doca-setup`](../doca-setup/SKILL.md)); or any hardware-state167change (use168[`doca-hardware-safety`](../doca-hardware-safety/SKILL.md)).169170## What this skill provides171172This is a **thin loader**. The substantive material lives in two173companion files:174175- `CAPABILITIES.md` — the collector deployment contract as a176 class: the four-surface decomposition (clx collection177 mechanism vs reader library vs publisher library vs productized178 DTS), the collection pipeline shape (providers / counters →179 schema → collector daemon → exporters), the export-backend180 surface (Prometheus pull, Fluent Bit push, NetFlow, file /181 IPC) at class level, the version overlay on182 [`doca-version`](../doca-version/SKILL.md), the error taxonomy183 (collector won't start → no provider rows → schema mismatch →184 exporter silent → downstream skew → transport), the185 observability surface, and the safety policy (gate provider186 support before commit; collector is read-only against the187 device; route any mutating step to188 [`doca-hardware-safety`](../doca-hardware-safety/SKILL.md); do189 not invent clx names / paths).190- `TASKS.md` — step-by-step workflows for the deployment verbs:191 `configure`, `build` (routing stub), `modify`, `run`, `test`,192 `debug`, plus a `Deferred task verbs` block that routes193 out-of-scope questions (the two libraries, the productized DTS194 container, env prep, hardware-state change) to their owners.195196The skill assumes a host or BlueField where DOCA is already197installed and healthy (per198[`doca-setup ## test`](../doca-setup/TASKS.md#test)) and the199operator can run the collector and reach its exporter sinks. It200does not cover installing DOCA — that path goes through201[`doca-setup`](../doca-setup/SKILL.md) — and it does not cover202the reader / publisher library APIs or the productized DTS203container.204205## Loading order2062071. Read this `SKILL.md` first to confirm the user's question is208 in scope (operating a CollectX-based collector and its209 exporters — NOT the reader / publisher library APIs, NOT the210 productized DTS container).2112. **For the four-surface decomposition, the collection pipeline212 shape, the export-backend class surface, the version overlay,213 the error taxonomy, the observability surface, and the safety214 policy, see [CAPABILITIES.md](CAPABILITIES.md).**2153. **For step-by-step workflows — `configure`, `build` (routing216 stub), `modify`, `run`, `test`, `debug`, and the217 `Deferred task verbs` block — see [TASKS.md](TASKS.md).**218219Both companion files cross-link to each other,220[`doca-version`](../doca-version/SKILL.md) for the canonical221version-handling rules, and222[`doca-public-knowledge-map`](../doca-public-knowledge-map/SKILL.md)223whenever the right answer is "read the live config / public docs"224rather than collector-specific guidance.225226## Example questions this skill answers well227228See [`references/details.md`](references/details.md#example-questions-this-skill-answers-well).229## What this skill deliberately does not ship230231See [`references/details.md`](references/details.md#what-this-skill-deliberately-does-not-ship).232## Related skills233234See [`references/details.md`](references/details.md#related-skills).