Version: 1.0.0
Last Updated: 2026-06-19
Applicability: Claude Code Skills
Core Standard: This skill implements Observability Standards. For the authoritative methodology (full Metrics/Traces detail, sampling, OTel integration), refer to the core standard.
Guide observability implementation across the three pillars: Logs, Metrics, and Traces.
引導三支柱可觀測性實作：Logs、Metrics、Traces。
Capabilities | 功能
Capability
Description
說明
Instrumentation Check
Pre-launch observability checklist
上線前可觀測性檢查表
Maturity Assessment
L0-L4 maturity self-evaluation
L0-L4 成熟度自評
Metric Design
Help design metrics (type, naming, labels)
協助設計 Metrics
Alert Design
Design SLO-based alerts with noise reduction
設計 SLO-based 告警
Golden Signals
Verify 4 golden signals coverage
驗證四大黃金信號覆蓋
Usage | 使用方式
/observability # Show observability guide
/observability --checklist # Run instrumentation checklist
/observability --maturity # Maturity assessment (L0-L4)
/observability --alerting # Alert design guide
/observability "payment-service"# Guide for specific service
Three Pillars Framework | 三支柱框架
Each pillar gives a different lens; their power is in correlation.
Pillar
What It Captures
When to Use
Granularity
Logs
Discrete events with context
Debugging, audit trails, error details
High (per-event)
Metrics
Numerical measurements over time
Dashboards, alerting, capacity planning
Low (aggregated)
Traces
Request flow across services
Latency analysis, dependency mapping
Medium (per-request)
Correlation fields: trace_id links Logs ↔ Traces ↔ Metrics (via Exemplars);
service.name filters all three pillars. Workflow: metric anomaly → exemplar →
trace → trace_id in logs.
Golden Signals | 四大黃金信號
Based on Google SRE. Every service SHOULD monitor all four before production.
Signal
Measure
Example Metric
Alert (SLO-based)
Latency
P50/P95/P99 via Histogram, split success/error
http.server.request.duration.seconds
P99 > X ms for 5 min
Traffic
Requests/sec, by route/method
http.server.request.total (rate)
drop > 50% or spike > 200%
Errors
errors / total requests
...request.total{status=~"5.."} ÷ total
error rate > X% for 5 min
Saturation
CPU/mem/pool/disk utilization
system.cpu.utilization (Gauge)
resource > 80% for 10 min
Metric Types | Metric 型別
Type
Behavior
Use When
Counter
Only goes up (resets on restart)
request count, error count, bytes sent
Gauge
Point-in-time, up/down
queue depth, active connections, memory
Histogram
Distribution across buckets
request duration, response size
Summary
Client-computed percentiles
legacy, no server-side aggregation
Naming: <domain>.<entity>.<action>.<unit> in snake_case (e.g.
db.client.query.duration.seconds). Label cardinality: keep labels under
~1000 unique values — never use user_id / request_id / raw url / ip as
labels; record those in Logs or Traces instead.
Maturity Model (L0–L4) | 成熟度模型
Level
Name
Characteristics
Upgrade Action
L0
No Observability
only stdout/stderr; debug via SSH + tail -f
structured logging; centralize collection
L1
Basic Logging
structured JSON logs, centralized, searchable
add business metrics; first dashboard
L2
Metrics-Driven
Logs + Metrics, dashboards, threshold alerts
enable tracing; SLO-based alerting
L3
Full Observability
three pillars + correlation + SLO alerts + Golden Signals
