Collectors Prometheus Profiles

Prometheus profile authoring

A profile is an operator dashboard design backed by source evidence, not a mechanical metric-name translation. The author decides the entity grain, chart comparisons, cardinality boundary, hierarchy, and collection policy; the validator proves only the contracts code can establish.

Authorities

Use these owners for format and runtime contracts. Verify affected enforcement claims against the current code and tests; the lists below are a source map, not evidence that a changed implementation still behaves that way:

• src/go/plugin/go.d/collector/prometheus/profile-format.md: the envelope, the runtime processing order, the stock contribution policy for autogen.selector and relabeling, chart-template rules, job-side profile selection.
• src/go/plugin/framework/charttpl/README.md: every group and chart field, the validation rules, engine-derived behavior.
• src/go/pkg/relabel/README.md: relabel actions, stage order, histogram and summary safety, profile precedence.
• src/go/tools/prometheus-profile-validation/README.md: the validator CLI, safe job policy, what PASS establishes (one finding code per objective check), the warning classes.
• src/go/internal/promprofile/README.md: framework boundary, authority model, compile and replay flow, support composition, the latest-testdata model.
• src/go/plugin/go.d/collector/prometheus/profile-proofs/README.md and src/go/tools/prometheus-profile-proof/README.md: proof artifact contract, evidence boundary, external testdata contract, the evidence-dirs and verify commands.
• docs/NIDL-Framework.md: the NIDL model (families, contexts, instances, dimensions, labels).

src/go/plugin/go.d/collector/prometheus/README.md is a symlink to a generated integration page and is not an authority.

Choose the workflow

Read the row for the work at hand, not the whole skill.

Loading this skill for review or explanation does not authorize capture, installation, database reset or generation. Apply the authoring criteria to the assigned change; execute operational procedures only within the actual request.

Authoring workflow

1. Establish the evidence boundary. Inventory metric families, types, labels, optional modes, and source revisions. Separate exposition facts (names, HELP, TYPE, labels, observed values), source facts (lifecycle, units, label domains, relationships, optionality), and design judgments (grain, composition, hierarchy, exclusions). A captured endpoint is private input: never commit it unless deliberately sanitized from public source contracts, and never commit credentials, customer identities, private endpoints, or deployment data. Missing source evidence is a real limitation; one observed fixture is not a universal exporter contract.
2. Design the operator model before YAML (chart-design.md). Answer, in causal order: is the service available and doing useful work; what load is it serving; is latency, error rate, or saturation worsening; which bounded entity or category explains it; which resource or dependency is responsible. For each view state one operator question, one entity grain, the smallest stable identity, the labels compared as dimensions, the labels kept for filtering, the labels omitted with the reducer that keeps the omission truthful, and the exact source signals with units, lifecycle, and inter-dimension relationship. Do not build an aggregate view Netdata derives by grouping the detailed one; choose the finest operator-useful grain whose cardinality and churn stay acceptable.
3. Classify labels and cardinality (chart-design.md, "Assign labels by role"; docs/NIDL-Framework.md when choosing a monitored component, instance grain, dimension set, or label role). Give every relevant label a role: required identity, optional identity, dimension, promoted metadata, routing only, or omitted with a stated lost comparison. Estimate cardinality from the exporter contract, not the fixture; raw user IDs, addresses, request IDs, URLs, exception text, and hashes are normally too high-cardinality for identity or dimensions.
4. Choose truthful aggregation (chart-design.md, "Aggregation when labels are omitted"). One reducer per chart; never merge gauges and counters into one rendered dimension, although a chart may hold distinct authored dimensions of different kinds when the shared comparison is intentional, each keeping its runtime-derived algorithm; quantiles are not mergeable. In a stock profile write aggregation explicitly wherever a deliberate many-to-one projection can occur, including deliberate sum, and omit it for collision-free routes.
5. Choose collection policy. match detects the exporter and bounds the profile's source namespace; prefer exporter-unique families for detection: generic runtime families (process_*, python_*, http_*) may still be charted, but naming them in match makes unrelated endpoints eligible. Profile relabeling normalizes or drops exporter-owned families after selection (recover bounded identity encoded in metric names, normalize label values such as status classes, remove an established useless family class such as a source-wide *_created, or another source-backed transformation), only with a source-backed contract and a bounded result, never as a substitute for an identity or aggregation decision; the first applicable selected profile owns each original family and every selected template consumes the shared result. Profile fallback_type classifies untyped scalars the exporter owns, as narrowly as the evidence allows. autogen.selector shapes only the generic fallback charts. Job selector, job relabeling, and job fallback_type are deployment policy and are never a hidden prerequisite of a stock profile. Unknown future families inside a wildcard namespace stay eligible for generic fallback; a bounded, source-proven alias may route a future input to an authored metric only when an explicit future_inputs case proves that branch.
6. Encode the profile. Nested groups express hierarchy: child context_namespace segments join with ., child family segments with /. Omit the root family when it would only repeat the resolved application; the named child groups then become top-level families, nested groups still need family, and a chart directly under a transparent root needs its own family. Keep a meaningful root on reusable instrumentation profiles that compose into other applications. Use base units; convert bytes to bits only where the operator convention is bandwidth. Omit algorithm (the runtime kind resolves it), type for line charts and histogram buckets (forced to heatmap), and multiplier and divisor defaults; use area and stacked only for the meanings in chart-design.md, "Choose chart types for visual meaning". Status values are dimensions from the source's closed state mapping, not chart instances. Gauge families ending in _info are skipped by the writer and never reach metrix (metric-types.md, "Info families").
7. Run the objective validator and the ToC ("Scripts" below). PASS proves schema and the exercised production collector, planner, and emitter path; it does not prove operator usefulness, source semantics, cardinality outside the evidence, or additivity. Resolve every warning with evidence; never silence one mechanically.
8. Review as an operator (chart-design.md, "Semantic review"). For a stock profile continue with ownership-proof.md and proof-authoring.md; stock work is complete only when the source contract, design, descriptor, fixtures, production profile, and integration metadata reconcile.

Stock contribution rule sheet

Stock profiles live under src/go/plugin/go.d/config/go.d/prometheus.profiles/default/<name>.yaml with a proof directory beside the collector. The "review by hand" table is contribution policy that the runtime format permits in user profiles. The validator has no user mode: its contributor-policy checks below apply to every profile it validates.

The code enforces

• Validator (tools/prometheus-profile-validation, finding codes in src/go/tools/prometheus-profile-validation/README.md#what-pass-establishes): no autogen.selector.allow; every deny names one exact family (open_ended_profile_fallback_deny), and under CLI validation that family must appear in the fixture (unproven_profile_fallback_deny; under proof replay the coverage discharge below covers it); current evidence yields zero generic fallback and zero unmatched series (unmatched series downgrade to the profile_suppressed_series warning when a selected profile's autogen.selector explains every one); the relabel grammar and name-provenance rules (under proof replay, three specific checks are deferred to warnings, one branch each of open_ended_relabel_name_rewrite, unpreserved_relabel_name_identity, and unbounded_relabel_discard; the other branches emitting the first two codes stay errors); a lifecycle cap that discards observed entities or dimensions; a chart with no visible dimension; bucket charts use observations/s and no algorithm other than incremental (omitting it is fine); explicit area or stacked raises a semantic-review warning; a selected series carrying a label the chart neither uses nor excludes raises a warning.
• Proof loaders and compiler (internal/promprofile/semantics): documentation.title and summary required; the closed evidence kind set; the outcome literals drop_before_writer and retain_writable_unrendered, with metadata_only bound to the latter; a reusable component, label, or reduction policy in either document needs two consumers (compile stage, semantics/evidence.go); every production autogen.selector.deny family is discharged by a retain_writable_unrendered exclusion naming it (coverage.go); metadata_only requires the conditions in metric-types.md, "Info families"; supports are accepted only in PROFILE-DESIGN.composition.supports; no other strict schema has the field.
• Integrations projection (integrations/prometheus_profile_docs.py, run by gen_integrations.py and integrations/tests/test_prometheus_profile_docs.py): every stock profile, supporting ones included, needs a direct row under some module in the top-level profile_coverage.modules of the Prometheus collector metadata.yaml (generation raises Stock Prometheus profiles without an integration mapping otherwise); the projection resolves the support closure from the design; the key is accepted only on go.d.plugin/prometheus modules (a Python check, not the JSON schema); an unknown module id under it is recorded as a warning by _common.py, and gen_integrations.py ends with fail_on_warnings(), so the generation run still fails; a semantic view and its runtime chart must agree on family and chart identity (also enforced by the proof compiler, semantics/replay_route.go); the view question is never rendered.