Metrics & SLOs (DOCS-OBS-51-001)

Audience: SREs and developers defining service-level objectives and wiring alerts for Stella Ops. Scope: Platform-wide metric families, suggested SLOs, alert examples, triage and TTFS KPIs, Offline Kit metrics, and observability hygiene.

Last updated: 2025-12-15

This guide is the platform-wide reference for what Stella Ops measures and the objectives those measurements serve. Module-specific signals (for example Aggregation Observability and FN-Drift) build on the conventions defined here.

Core metrics (platform-wide)

• Requests: http_requests_total{tenant,workload,route,status} (counter); latency histogram http_request_duration_seconds.
• Jobs: worker_jobs_total{tenant,queue,status}; worker_job_duration_seconds.
• DB: db_query_duration_seconds{db,operation}; db_pool_in_use, db_pool_available.
• Cache: cache_requests_total{result=hit|miss}; cache_latency_seconds.
• Queue depth: queue_depth{tenant,queue} (gauge).
• Errors: errors_total{tenant,workload,code}.
• Custom module metrics: keep namespaced (e.g., riskengine_score_duration_seconds, notify_delivery_attempts_total).

SLOs (suggested)

• API availability: 99.9% monthly per public service.
• P95 latency: <300 ms for read endpoints; <1 s for write endpoints.
• Worker job success: >99% over 30d; P95 job duration set per queue (document locally).
• Queue backlog: alert when queue_depth > 1000 for 5 minutes per tenant/queue.
• Error budget policy: 28-day rolling window; burn-rate alerts at 2× and 14× budget.

Alert examples

• High error rate: rate(errors_total[5m]) / rate(http_requests_total[5m]) > 0.02.
• Latency regression: histogram_quantile(0.95, sum(rate(http_request_duration_seconds_bucket[5m])) by (le,route)) > 0.3.
• Queue backlog: queue_depth > 1000 for 5m.
• Job failures: rate(worker_jobs_total{status="failed"}[10m]) > 0.01.

UX KPIs (triage TTFS)

• Targets:
  • TTFS first evidence p95: <= 1.5s
  • TTFS skeleton p95: <= 0.2s
  • Clicks-to-closure median: <= 6
  • Evidence completeness avg: >= 90% (>= 3.6/4)

# TTFS first evidence p50/p95
histogram_quantile(0.50, sum(rate(stellaops_ttfs_first_evidence_seconds_bucket[5m])) by (le))
histogram_quantile(0.95, sum(rate(stellaops_ttfs_first_evidence_seconds_bucket[5m])) by (le))

# Clicks-to-closure median
histogram_quantile(0.50, sum(rate(stellaops_clicks_to_closure_bucket[5m])) by (le))

# Evidence completeness average percent (0-4 mapped to 0-100)
100 * (sum(rate(stellaops_evidence_completeness_score_sum[5m])) / clamp_min(sum(rate(stellaops_evidence_completeness_score_count[5m])), 1)) / 4

# Budget violations by phase
sum(rate(stellaops_performance_budget_violations_total[5m])) by (phase)

• Dashboard: docs/modules/telemetry/operations/dashboards/ttfs-observability.json
• Alerts: devops/telemetry/alerts/triage-alerts.yaml

TTFS Metrics (time-to-first-signal)

• Core metrics:

  • ttfs_latency_seconds{surface,cache_hit,signal_source,kind,phase,tenant_id} (histogram)
  • ttfs_signal_total{surface,cache_hit,signal_source,kind,phase,tenant_id} (counter)
  • ttfs_cache_hit_total{surface,cache_hit,signal_source,kind,phase,tenant_id} (counter)
  • ttfs_cache_miss_total{surface,cache_hit,signal_source,kind,phase,tenant_id} (counter)
  • ttfs_slo_breach_total{surface,cache_hit,signal_source,kind,phase,tenant_id} (counter)
  • ttfs_error_total{surface,cache_hit,signal_source,kind,phase,tenant_id,error_type,error_code} (counter)

• SLO targets:

  • P50 < 2s, P95 < 5s (all surfaces)
  • Warm path P50 < 700ms, P95 < 2.5s
  • Cold path P95 < 4s

# TTFS latency p50/p95
histogram_quantile(0.50, sum(rate(ttfs_latency_seconds_bucket[5m])) by (le))
histogram_quantile(0.95, sum(rate(ttfs_latency_seconds_bucket[5m])) by (le))

# SLO breach rate (per minute)
60 * sum(rate(ttfs_slo_breach_total[5m]))

Offline Kit (air-gap) metrics

• offlinekit_import_total{status,tenant_id} (counter)
• offlinekit_attestation_verify_latency_seconds{attestation_type,success} (histogram)
• attestor_rekor_success_total{mode} (counter)
• attestor_rekor_retry_total{reason} (counter)
• rekor_inclusion_latency{success} (histogram)

# Import rate by status
sum(rate(offlinekit_import_total[5m])) by (status)

# Import success rate
sum(rate(offlinekit_import_total{status="success"}[5m])) / clamp_min(sum(rate(offlinekit_import_total[5m])), 1)

# Attestation verify p95 by type (success only)
histogram_quantile(0.95, sum(rate(offlinekit_attestation_verify_latency_seconds_bucket{success="true"}[5m])) by (le, attestation_type))

# Rekor inclusion latency p95 (by success)
histogram_quantile(0.95, sum(rate(rekor_inclusion_latency_bucket[5m])) by (le, success))

Dashboard: docs/modules/telemetry/dashboards/offline-kit-operations.json

Observability hygiene

• Tag everything with tenant, workload, env, region, version.
• Keep metric names stable; prefer adding labels over renaming.
• No high-cardinality labels (avoid user_id, path, raw errors); bucket or hash if needed.
• Offline: scrape locally (Prometheus/OTLP); ship exports via bundle if required.

Dashboards

• Golden signals per service: traffic, errors, saturation, latency (P50/P95/P99).
• Queue dashboards: depth, age, throughput, success/fail rates.
• Tracing overlays: link span status to error metrics; use exemplars where supported.

Validation checklist

• [ ] Metrics emitted with required tags.
• [ ] Cardinality review completed (no unbounded labels).
• [ ] Alerts wired to error budget policy.
• [ ] Dashboards cover golden signals and queue health.