Agent Skills: Monitoring Setup

Monitoring Setup

Position in workflow: v0.8 Runbook Creation → v0.8 Monitoring Setup → v0.9 GTM Strategy

Consumes

This skill requires prior work from v0.8 Runbook Creation and earlier stages:

• RUN-* runbook entries (from v0.8 Runbook Creation) — Incident response runbooks define alerting scenarios; critical alerts must link to RUN- procedures
• DEP-* deployment entries (from v0.8 Release Planning) — DEP- rollback thresholds and post-deploy validation steps inform MON- alert conditions and SLO targets
• API-* endpoint contracts (from v0.6 Technical Specification) — Define baseline latency, throughput, and error rates for application-layer metrics
• KPI-* metrics (from v0.3 Outcome Definition and v0.9 Launch Metrics) — Business metrics (signups, conversions, retention) inform dashboard design and business layer monitoring
• ARC-* architecture decisions (from v0.6 Architecture Design) — System structure determines which components to monitor (monolith has different metrics than distributed services)
• TECH-* technology stack (from v0.5 Technical Stack Selection) — Technology choices (database, cloud provider, APM tools) determine available metrics and monitoring tools

This skill assumes DEP- and RUN- entries are complete with thresholds, rollback conditions, and incident procedures defined.

Produces

This skill creates/updates:

• MON-* entries (monitoring specifications, metric/alert/dashboard/SLO types) — Concrete monitoring rules with thresholds, alert conditions, dashboards, SLO definitions, linked to RUN- procedures
• Alert routing configuration — Mapping of MON- alerts to notification channels and teams; links alerts to RUN- incident procedures
• Observability baseline — Metrics gathered from staging/production, establishing normal operating ranges for alert thresholds

All MON- entries are operational monitoring specifications, not confidence-based. They are:

• Measurable (every metric has a source, unit, and aggregation method)
• Actionable (every alert has a RUN- procedure; no orphaned alerts)
• Thresholded (critical/warning severity with specific numeric conditions)
• Dashboarded (MON- dashboard entries provide visibility to operators and stakeholders)
• SLO-backed (SLO entries tie monitoring to product commitments)

Example MON- entries:

MON-001: API Request Latency (p95)
Type: Metric
Layer: Application
Owner: Backend Team

Name: api.request.latency.p95
Description: 95th percentile response time for all API endpoints (from API-001–020)
Unit: ms
Source: Application APM (Datadog custom instrumentation)
Aggregation: p95 over 5-minute window
Retention: 90 days

Linked IDs: API-001 to API-020, DEP-004 (baseline from staging)

---

MON-002: High Latency Alert (Warning)
Type: Alert
Layer: Application
Owner: Backend Team

Metric: MON-001 (api.request.latency.p95)
Condition: >500ms (from DEP-002 baseline)
Window: 5 minutes
Severity: Warning
Runbook: RUN-001 (Performance Degradation Investigation)

Notification:
  - Channel: Slack #backend-alerts
  - Recipients: Backend on-call, team notified during business hours

Silencing: During scheduled maintenance windows (DEP-004 notifications)

Linked IDs: MON-001, RUN-001, DEP-002

---

MON-003: Critical Latency Alert
Type: Alert
Layer: Application
Owner: Backend Team

Metric: MON-001 (api.request.latency.p95)
Condition: >2000ms (SLA breach, from KPI-001 target)
Window: 2 minutes
Severity: Critical
Runbook: RUN-001 (Performance Degradation Investigation)

Notification:
  - Channel: PagerDuty (wake on-call)
  - Recipients: Backend on-call, Tech Lead, escalate if not acknowledged in 5 min

Silencing: None (critical alerts never silenced)

Linked IDs: MON-001, RUN-001, KPI-001

---

MON-004: API Availability SLO
Type: SLO
Layer: Application
Owner: Platform Team

Objective: API endpoints return non-5xx response
Target: 99.9% uptime (from DEP-002 / KPI-001)
Window: Rolling 30 days
Error Budget: 43.2 minutes/month

Alerting:
  - 50% error budget consumed → Warning to engineering (slow-burn alert)
  - 75% error budget consumed → Critical, freeze non-essential deploys
  - 100% error budget consumed → Post-incident review required (RUN-008 procedure)

Linked IDs: API-001–020, DEP-003 (rollback triggers), RUN-008 (incident review)

---

MON-005: System Health Dashboard
Type: Dashboard
Layer: Infrastructure + Application
Owner: Platform Team

Purpose: Quick health check for on-call engineers (run from RUN-002, RUN-001)
Audience: On-call engineers, engineering leadership, ops team
Panels:
  - API Request Rate (last 1h): Should be steady or increasing
  - API Latency (p50, p95, p99): Watch for p95/p99 creeping up
  - Error Rate by Endpoint: Any 5xx > 0 is concerning
  - Active Critical Alerts: Should be none
  - Database Connection Pool (from MON-006): Trending toward threshold
  - CPU/Memory by Service: Identify resource exhaustion
  - Deployment Status: Current version, time of last deploy
Refresh: 30 seconds

Linked IDs: MON-001, MON-002, MON-003, MON-006, DEP-001, RUN-001/002

---

MON-006: Database Connection Pool Utilization
Type: Metric
Layer: Infrastructure
Owner: Database Team

Name: db.connection_pool.utilized_percent
Description: Percentage of available connections in use (from DEP-001 pool size)
Unit: percentage
Source: Database monitoring (RDS Enhanced Monitoring or custom query)
Aggregation: avg over 1-minute window
Retention: 30 days

Linked IDs: DEP-001 (pool config), RUN-001 (incident when >90%)

Core Concept: Monitoring as Early Warning

Monitoring is not about collecting data—it is about detecting problems before users do. Every metric should answer: "Is this working? If not, what's broken?"

Monitoring Layers

| Layer | What to Measure | Why It Matters | |-------|-----------------|----------------| | Infrastructure | CPU, memory, disk, network | System health foundation | | Application | Latency, errors, throughput | User-facing performance | | Business | Signups, conversions, revenue | Product health | | User Experience | Page load, interaction time | Real user impact |

Execution

1. Define SLOs (Service Level Objectives)

   • What uptime do we promise?
   • What latency is acceptable?
   • What error rate is tolerable?

2. Identify key metrics per layer

   • Infrastructure: Resource utilization
   • Application: RED metrics (Rate, Errors, Duration)
   • Business: KPI- from v0.3 and v0.9
   • User: Core Web Vitals, journey completion

3. Set alert thresholds

   • Warning: Investigate soon
   • Critical: Act immediately
   • Base on SLOs and historical data

4. Map alerts to runbooks

   • Every critical alert → RUN- procedure
   • No alert without action path

5. Design dashboards

   • Overview: System health at a glance
   • Deep-dive: Per-service details
   • Business: KPI tracking

6. Create MON- entries with full traceability

MON- Output Template

MON-XXX: [Monitoring Rule Title]
Type: [Metric | Alert | Dashboard | SLO]
Layer: [Infrastructure | Application | Business | User Experience]
Owner: [Team responsible for this metric/alert]

For Metric Type:
  Name: [metric.name.format]
  Description: [What this measures]
  Unit: [count | ms | percentage | bytes]
  Source: [Where this comes from]
  Aggregation: [avg | sum | p50 | p95 | p99]
  Retention: [How long to keep data]

For Alert Type:
  Metric: [MON-YYY or metric name]
  Condition: [Threshold expression]
  Window: [Time window for evaluation]
  Severity: [Critical | Warning | Info]
  Runbook: [RUN-XXX to follow when fired]
  Notification:
    - Channel: [Slack, PagerDuty, Email]
    - Recipients: [Team or individuals]
  Silencing: [When to suppress, e.g., maintenance windows]

For Dashboard Type:
  Purpose: [What questions this answers]
  Audience: [Who uses this dashboard]
  Panels: [List of visualizations]
  Refresh: [How often to update]

For SLO Type:
  Objective: [What we promise]
  Target: [Percentage, e.g., 99.9%]
  Window: [Rolling 30 days]
  Error Budget: [How much downtime allowed]
  Alerting: [When error budget is at risk]

Linked IDs: [API-XXX, UJ-XXX, KPI-XXX, RUN-XXX related]

Example MON- entries:

MON-001: API Request Latency (p95)
Type: Metric
Layer: Application
Owner: Backend Team

Name: api.request.latency.p95
Description: 95th percentile response time for all API endpoints
Unit: ms
Source: Application APM (Datadog/New Relic)
Aggregation: p95
Retention: 90 days

Linked IDs: API-001 to API-020

MON-002: High Latency Alert
Type: Alert
Layer: Application
Owner: Backend Team

Metric: MON-001 (api.request.latency.p95)
Condition: > 500ms
Window: 5 minutes
Severity: Warning
Runbook: RUN-006 (Performance Degradation Investigation)

Notification:
  - Channel: Slack #backend-alerts
  - Recipients: Backend on-call

Silencing: During scheduled deployments (DEP-002 windows)

Linked IDs: MON-001, RUN-006, DEP-002

MON-003: Critical Latency Alert
Type: Alert
Layer: Application
Owner: Backend Team

Metric: MON-001 (api.request.latency.p95)
Condition: > 2000ms
Window: 2 minutes
Severity: Critical
Runbook: RUN-006 (Performance Degradation Investigation)

Notification:
  - Channel: PagerDuty
  - Recipients: Backend on-call, Tech Lead

Silencing: None (always alert on critical)

Linked IDs: MON-001, RUN-006

MON-004: API Availability SLO
Type: SLO
Layer: Application
Owner: Platform Team

Objective: API endpoints return non-5xx response
Target: 99.9%
Window: Rolling 30 days
Error Budget: 43.2 minutes/month

Alerting:
  - 50% budget consumed → Warning to engineering
  - 75% budget consumed → Critical, freeze non-essential deploys
  - 100% budget consumed → Incident review required

Linked IDs: API-001 to API-020, DEP-003

MON-005: System Health Dashboard
Type: Dashboard
Layer: Infrastructure + Application
Owner: Platform Team

Purpose: Quick health check for on-call engineers
Audience: On-call, engineering leadership
Panels:
  - API Request Rate (last 1h)
  - API Latency (p50, p95, p99)
  - Error Rate by Endpoint
  - Active Alerts
  - Database Connection Pool
  - CPU/Memory by Service
Refresh: 30 seconds

Linked IDs: MON-001, MON-002, MON-003

The RED Method (Application Monitoring)

For each service, measure:

| Metric | What It Measures | Alert Threshold | |--------|------------------|-----------------| | Rate | Requests per second | Anomaly detection | | Errors | Failed requests / total | >1% warning, >5% critical | | Duration | Request latency (p95, p99) | >500ms warning, >2s critical |

The USE Method (Infrastructure Monitoring)

For each resource (CPU, memory, disk, network):

| Metric | What It Measures | Alert Threshold | |--------|------------------|-----------------| | Utilization | % of capacity used | >80% warning, >95% critical | | Saturation | Queue depth, waiting | >0 for critical resources | | Errors | Error count/rate | Any errors = investigate |

SLO Framework

| Tier | Availability | Latency (p95) | Use For | |------|--------------|---------------|---------| | Tier 1 | 99.99% (52 min/yr) | <100ms | Payment, auth | | Tier 2 | 99.9% (8.7 hr/yr) | <500ms | Core features | | Tier 3 | 99% (3.6 days/yr) | <2s | Background jobs |

Alert Severity Matrix

| Severity | User Impact | Response Time | Notification | |----------|-------------|---------------|--------------| | Critical | Service unusable | <5 min | PagerDuty (wake up) | | Warning | Degraded experience | <30 min | Slack (business hours) | | Info | No immediate impact | Next day | Dashboard/log |

Dashboard Design Principles

| Principle | Implementation | |-----------|----------------| | Answer questions | Each panel answers "Is X working?" | | Hierarchy | Overview → Service → Component | | Context | Show thresholds, comparisons | | Actionable | Link to runbooks from alerts | | Fast | Quick load, auto-refresh |

Anti-Patterns

| Pattern | Signal | Fix | |---------|--------|-----| | Alert fatigue | Too many alerts, team ignores | Tune thresholds, remove noise | | No runbook link | Alert fires, no one knows what to do | Every alert → RUN- | | Vanity metrics | "1 million requests!" without context | Focus on user-impacting metrics | | Missing baselines | No historical comparison | Establish baselines before launch | | Over-monitoring | 500 metrics, can't find signal | Focus on RED/USE fundamentals | | Under-monitoring | "We'll add monitoring later" | Monitoring ships with code |

Quality Gates

Before proceeding to v0.9 GTM Strategy:

• [ ] SLOs defined for critical services (MON- SLO type)
• [ ] RED metrics configured for application layer
• [ ] USE metrics configured for infrastructure layer
• [ ] Critical alerts linked to RUN- procedures
• [ ] Overview dashboard created for on-call
• [ ] Alert notification channels configured
• [ ] Baseline metrics established from staging

Downstream Connections

| Consumer | What It Uses | Example | |----------|--------------|---------| | On-Call Team | MON- alerts trigger response | MON-003 → page engineer | | v0.9 Launch Metrics | MON- provides baseline data | MON-001 baseline → KPI-010 target | | Post-Mortems | MON- data for incident analysis | "MON-005 showed spike at 14:32" | | Capacity Planning | MON- trends inform scaling | USE metrics → infrastructure planning | | DEP- Rollback | MON- thresholds trigger rollback | MON-002 breach → DEP-003 rollback |

Detailed References

• Monitoring stack examples: See references/monitoring-stack.md
• MON- entry template: See assets/mon-template.md
• SLO calculation guide: See references/slo-guide.md
• Dashboard best practices: See references/dashboard-guide.md