Description
An NVL72 rack is one NVLink domain of 72 GPUs joined through NVSwitch trays. NMX-Telemetry collects from the domain over gNMI and pushes OTLP from its built-in exporter, NMX-Controller manages partitions, and NetQ NVLink provides the domain and partition view and can re-export OTLP. The degraded-domain journey needs domain state, partition state, NVSwitch health, per-port link state and errors, and recovery activity, alongside NMX logs. Per-port telemetry is one of the largest cardinality drivers in these estates, and the port count itself has to be confirmed rather than assumed.
Acceptance Criteria
- #1 The construct emits NMX-Telemetry OTLP metrics and NetQ NVLink domain and partition state with names from the NVIDIA documentation at a pinned release, in a new signals/ area file
- #2 Per-port series count for one NVL72 rack is documented with its basis and visible in just dump
- #3 NVLink port errors on a GPU line up with the DCGM NVLink counters for the same GPU identity
- #4 NMX logs and events are emitted in the shape NMX actually ships (OTLP logs or syslog through the shared mechanic)
- #5 Failure modes cover a degraded domain, an NVSwitch tray failure and an NVLink link-error burst
- #6 Every metric, label, attribute and log field is sourced from vendor documentation or a capture at a pinned version and recorded with provenance in the named signals/ area file; anything unconfirmed is a PENDING in cantfind.md, never an invented name
- #7 Rates, counts and cardinality in any shipped blueprint are plausible for a real deployment, with the real-world basis stated in a comment; nothing is raised to force emission
Definition of Done
- #1 just check (fmt-check, lint, gen-check, env-check, docs-check, test, race, hygiene, ui-check, compose-check, helm-test, lab-check, signal-fidelity)
- #2 just gen (only if a blueprint field, construct/workload config struct, or a skill under plugins/synthkit/skills/ changed)
- #3 just dump — inventory diffed against signals/
Implementation Notes
loop47: admitted R09 NVLink/NMX vendor-contract and physical endpoint mapping research; implementation waits for fixture/syslog/DCGM landing and source-confirmed map
loop47: two research passes found current NMX-T 1.2.3 docs; full metrics catalogue is NVOnline ID 1092300 or documented management/schema response, unavailable here. Physical GPU-link-to-port/DCGM mapping and NetQ NVLink exporter contract remain missing. Resume with those pinned artifacts/captures; fail closed, no invented fields or port counts. Implementation attempts 0; review-repairs 0; infra retries 0; not rejected.