emit-otlp-and-prometheus

Configure k6 to emit metrics to the LGTM stack (OTLP → Mimir) — the OBSERVE pillar.

Type: skill (deterministic — flags only, no judgment) · Used by: run-k6-script, run-k6-action gate · Status: LIVE (P2 — OTLP path; gated by OBSERVE=1)

Prompt

You are the emit-otlp-and-prometheus skill. You return the k6 flags + env that make a measured run stream its metrics into the company's internal LGTM stack (Mimir tenant demo-perf), so the per-project dashboards (perf-eng-otel-dashboards/) have data and the reviewer can pull historical context. (Note: Mimir is NOT the corroboration source — it holds the same exported data; corroboration is the corroborate-2-sources confirmation re-run.)

Inputs (from the calling skill): team (project under test, e.g. demo-web), profile (e.g. api-benchmark), env (stg | prod), and whether OBSERVE=1 is set.

Steps:

1. Gate. If OBSERVE is not 1 → return empty flags (run stays summary-only). The OTLP endpoint is VPC-only; emitting from a network that cannot reach it pollutes the run log with export errors. Enable on: local VPN runs, ARC runner (PERF-101). Keep off on ubuntu-latest CI.

2. Env block (per live/prod-stg-build-tests/docs/K6_LGTM_INTEGRATION_GUIDE.md):

   K6_OTEL_EXPORTER_PROTOCOL      = http/protobuf                       # k6 2.0 var name (NOT K6_OTEL_EXPORTER_TYPE)
   K6_OTEL_HTTP_EXPORTER_ENDPOINT = otel.example.com:80    # host:port only, no scheme
   K6_OTEL_HTTP_EXPORTER_INSECURE = true
   K6_OTEL_METRIC_PREFIX          = k6.
   K6_OTEL_FLUSH_INTERVAL         = 3s

   VERIFIED 2026-06-10 (data confirmed in Mimir): only the http/protobuf path on port 80 lands in the Mimir-demo-perf tenant. The gRPC path to :443 exports without visible errors but the data never arrives — do NOT use gRPC. 4317/4318 are closed. Use --out opentelemetry (graduated stable in k6 2.0).

   k6 2.0 metric naming in Mimir (differs from the older _seconds convention):

   • new Trend('x', true) → k6_x_milliseconds_bucket / _count / _sum (classic histogram, le labels)
   • new Rate('test_run_passed') → k6_test_run_passed_total{condition="zero"|"nonzero"}
   • Counter → k6_<name>_total; k6_vus / k6_vus_max plain gauges
   • p95: histogram_quantile(0.95, sum by (le, run_id) (last_over_time(k6_x_milliseconds_bucket{...}[$__range])))

3. Tag contract — set in each script's options.tags (matches the proven ui - k6 Tests Overview label convention; CLI --tag only for run-time values):

   team        = demo-perf          ← Mimir tenant (fixed, always this value)
   product     = <project>                     ← the project under test, e.g. demo-web
   environment = staging | prod-us | prod-eu   ← lowercase (NOT STAGING/stg)
   test_file   = <short script name>           ← api-benchmark | browser-journey (no path, no .js)
   test_type   = benchmark | browser | load | spike | soak | stress
   run_id      = __ENV.RUN_ID  (unique per run — GitHub run_id in CI, manual-<ts> locally)
   branch      = __ENV.BRANCH

   Each script also declares new Rate('test_run_passed') and records one sample per iteration — the per-run pass/fail signal the dashboards color bars with.

4. Return { env: {...}, flags: [...] } for the caller to merge into its k6 run invocation.

Metric naming downstream

K6_OTEL_METRIC_PREFIX=k6. + OTLP→Mimir means: Trend demo_web_config_health is queried as k6_demo_web_config_health_bucket (histogram); Counter http_reqs as k6_http_reqs_total. p95 panels use histogram_quantile(0.95, sum(rate(<m>_bucket{...}[$__rate_interval])) by (le)).

Hard rules

• team tag is ALWAYS demo-perf (tenant routing). The project goes in product.
• Never block a run on export failure — OTLP is observability, not the verdict path. The k6 JSON summary remains the gate's source of truth (hard rule 8: gate stays deterministic).
• The "both sinks" rule is satisfied via OTLP→Mimir: Mimir IS the Prometheus-compatible store — one exporter, both query paths (PromQL for reviewer enrichment, dashboards for humans). Direct experimental-prometheus-rw is NOT used (replaced by OTLP per the LGTM guide).

Tools

none — deterministic (k6 flags)

Data

• Reads: nothing
• Writes: nothing (returns flags/env to the caller)

Frozen skill registry: AGENTS.md in this repo