observability-design（可观察性设计）

Observability Patterns

When to Use

• Designing monitoring infrastructure for new services
• Defining SLIs, SLOs, and error budgets for reliability
• Implementing distributed tracing across microservices
• Creating alert rules that minimize noise and maximize signal
• Building dashboards for operations and business stakeholders
• Establishing incident response and postmortem processes
• Diagnosing production issues through telemetry correlation

Philosophy

You cannot fix what you cannot see. Observability is not about collecting data - it is about answering questions you have not thought to ask yet. Good observability turns every incident into a learning opportunity and every metric into actionable insight.

The Three Pillars

Metrics

Numeric measurements aggregated over time. Best for understanding system behavior at scale.

Characteristics:

• Highly efficient storage (aggregated values)
• Support mathematical operations (rates, percentiles)
• Enable alerting on thresholds
• Limited cardinality (avoid high-cardinality labels)

Types:

Logs

Immutable records of discrete events. Best for understanding specific occurrences.

Characteristics:

• Rich context and arbitrary data
• Expensive to store and query at scale
• Essential for debugging specific issues
• Should be structured (JSON) for parseability

Structure:

Required fields:
- timestamp: ISO 8601 format with timezone
- level: ERROR, WARN, INFO, DEBUG
- message: Human-readable description
- service: Service identifier
- trace_id: Correlation identifier

Context fields:
- user_id: Sanitized user identifier
- request_id: Request correlation
- duration_ms: Operation timing
- error_type: Classification for errors

Traces

Records of request flow across distributed systems. Best for understanding causality and latency.

Characteristics:

• Show request path through services
• Identify latency bottlenecks
• Reveal dependencies and failure points
• Higher overhead than metrics

Components:

• Trace: Complete request journey
• Span: Single operation within a trace
• Context: Metadata propagated across services

SLI/SLO/SLA Framework

Service Level Indicators (SLIs)

Quantitative measures of service behavior from the user perspective.

Common SLI categories:

SLI specification:

SLI: Request Latency
Definition: Time from request received to response sent
Measurement: Server-side histogram at p50, p95, p99
Exclusions: Health checks, internal tooling
Data source: Application metrics

Service Level Objectives (SLOs)

Target reliability levels for SLIs over a time window.

SLO formula:

SLO = (Good events / Total events) >= Target over Window

Example:
99.9% of requests complete successfully in < 200ms
measured over a 30-day rolling window

Setting SLO targets:

• Start with current baseline performance
• Consider user expectations and business impact
• Balance reliability investment against feature velocity
• Document error budget policy

Error Budgets

The allowed amount of unreliability within an SLO.

Calculation:

Error Budget = 1 - SLO Target

99.9% SLO = 0.1% error budget
= 43.2 minutes downtime per 30 days
= 8.64 seconds per day

Error budget policies:

• Budget remaining: Continue feature development
• Budget depleted: Focus on reliability work
• Budget burning fast: Freeze deploys, investigate

Alerting Strategies

Symptom-Based Alerts

Alert on user-visible symptoms, not internal causes.

Good alerts:

• Error rate exceeds threshold (users experiencing failures)
• Latency SLO at risk (users experiencing slowness)
• Queue depth growing (backlog affecting users)

Poor alerts:

• CPU at 80% (may not affect users)
• Pod restarted (self-healing, may not affect users)
• Disk at 70% (not yet impacting service)

Multi-Window, Multi-Burn-Rate Alerts

Detect fast burns quickly, slow burns before budget depletion.

Configuration:

Fast burn: 14.4x burn rate over 1 hour
  - Fires in 1 hour if issue persists
  - Catches severe incidents quickly

Slow burn: 3x burn rate over 3 days
  - Fires before 30-day budget depletes
  - Catches gradual degradation

Alert Fatigue Prevention

Strategies:

• Alert only on actionable issues
• Consolidate related alerts
• Set meaningful thresholds (not arbitrary)
• Require sustained condition before firing
• Include runbook links in every alert
• Review and prune alerts quarterly

Alert quality checklist:

• Can someone take action right now?
• Is the severity appropriate?
• Does it include enough context?
• Is there a runbook linked?
• Has it fired false positives recently?

Dashboard Design

Hierarchy of Dashboards

Service Health Overview:

• High-level SLO status
• Error budget consumption
• Key business metrics
• Designed for quick triage

Deep-Dive Diagnostic:

• Detailed metrics breakdown
• Resource utilization
• Dependency health
• Designed for investigation

Business Metrics:

• User-facing KPIs
• Conversion and engagement
• Revenue impact
• Designed for stakeholders

Dashboard Principles

• Answer specific questions, not show all data
• Use consistent color coding (green=good, red=bad)
• Show time ranges appropriate to the metric
• Include context (deployments, incidents) on graphs
• Mobile-responsive for on-call use
• Provide drill-down paths to detailed views

Essential Panels

Best Practices

• Correlate metrics, logs, and traces with shared identifiers
• Instrument code at service boundaries, not everywhere
• Use structured logging with consistent field names
• Set retention policies appropriate to data value
• Test alerts in staging before production
• Document SLOs and share with stakeholders
• Conduct regular game days to validate observability
• Automate common diagnostic procedures in runbooks

Anti-Patterns

• Alert on every possible metric (alert fatigue)
• Create dashboards without specific questions in mind
• Log without structure or correlation IDs
• Set SLOs without measuring current baseline
• Ignore error budget policies when convenient
• Treat all alerts with equal severity
• Store high-cardinality data in metrics (use logs/traces)
• Skip postmortems when issues resolve themselves

References

• references/monitoring-patterns.md - Detailed implementation patterns and examples