---

name: plan-capacity locale: caveman-ultra source_locale: en source_commit: 82c77053 fence_basis_commit: 82c77053 translator: "Julius Brussee homage — caveman" translation_date: "2026-04-26" description: > Capacity planning w/ historical metrics + growth models. Use predict_linear for forecast, find resource constraints, calc headroom, recommend scaling before saturation. Use → before traffic spikes/launches, quarterly reviews, rising util trends, budget cycles. license: MIT allowed-tools: Read Write Edit Bash Grep Glob metadata: author: Philipp Thoss version: "1.0" domain: observability complexity: intermediate language: multi tags: capacity-planning, forecasting, predict-linear, growth, headroom

Plan Capacity

Forecast resource needs, prevent saturation via data-driven capacity planning.

Use When

• Before seasonal traffic spikes (holidays, sales)
• New feature launches
• Quarterly capacity reviews
• Resource util trends up
• Before budget cycles

In

• Required: Historical metrics (CPU, mem, disk, net, req/s)
• Required: Time range for trend (min 4 weeks)
• Optional: Business growth projections (user growth, launches)
• Optional: Budget constraints

Do

Step 1: Collect Historical Metrics

Query Prometheus for key resources:

# CPU usage trend over 8 weeks
avg(rate(node_cpu_seconds_total{mode!="idle"}[5m])) by (instance)

# Memory usage trend
avg(node_memory_MemTotal_bytes - node_memory_MemAvailable_bytes) by (instance)

# Disk usage growth
avg(node_filesystem_size_bytes - node_filesystem_free_bytes) by (instance, device)

# Request rate growth
sum(rate(http_requests_total[5m])) by (service)

# Database connection pool usage
avg(db_connection_pool_used / db_connection_pool_max) by (instance)

Export → analyze:

# Export 8 weeks of CPU data
curl -G 'http://prometheus:9090/api/v1/query_range' \
  --data-urlencode 'query=avg(rate(node_cpu_seconds_total{mode!="idle"}[5m])) by (instance)' \
  --data-urlencode 'start=2024-12-15T00:00:00Z' \
  --data-urlencode 'end=2025-02-09T00:00:00Z' \
  --data-urlencode 'step=1h' | jq '.data.result' > cpu_8weeks.json

→ Clean time series per resource, no big gaps.

If err: missing data → forecast accuracy down. Check metric retention + scrape intervals.

Step 2: Calc Growth w/ predict_linear

Prometheus predict_linear() → forecast saturation:

# Predict when CPU will hit 80% (4 weeks ahead)
predict_linear(
  avg(rate(node_cpu_seconds_total{mode!="idle"}[5m]))[8w:],
  4*7*24*3600  # 4 weeks in seconds
) > 0.80

# Predict disk full date (8 weeks ahead)
predict_linear(
  avg(node_filesystem_size_bytes - node_filesystem_free_bytes)[8w:],
  8*7*24*3600
) > 0.95 * avg(node_filesystem_size_bytes)

# Predict memory pressure (2 weeks ahead)
predict_linear(
  avg(node_memory_MemTotal_bytes - node_memory_MemAvailable_bytes)[8w:],
  2*7*24*3600
) / avg(node_memory_MemTotal_bytes) > 0.90

# Predict request rate capacity breach (4 weeks ahead)
predict_linear(
  sum(rate(http_requests_total[5m]))[8w:],
  4*7*24*3600
) > 10000  # known capacity limit

Forecast dashboard:

{
  "dashboard": {
    "title": "Capacity Forecast",
    "panels": [
      {
        "title": "CPU Saturation Forecast (4 weeks)",
        "targets": [
          {
            "expr": "predict_linear(avg(rate(node_cpu_seconds_total{mode!=\"idle\"}[5m]))[8w:], 4*7*24*3600)",
            "legendFormat": "Predicted CPU"
          },
          {
            "expr": "0.80",
            "legendFormat": "Target Threshold (80%)"
          }
        ]
      },
      {
        "title": "Disk Full Date",
        "targets": [
          {
            "expr": "(avg(node_filesystem_size_bytes) - predict_linear(avg(node_filesystem_free_bytes)[8w:], 8*7*24*3600)) / avg(node_filesystem_size_bytes)",
            "legendFormat": "Predicted Usage %"
          }
        ]
      }
    ]
  }
}

→ Clear viz → when resources breach thresholds.

If err: predictions wrong (negative, wild swings) → check:

• Insufficient history (need 4+ weeks)
• Step spikes (deploys, migrations) distorting trend
• Seasonal patterns not in linear model

Step 3: Calc Current Headroom

Safety margin before saturation:

# CPU headroom (percentage remaining before 80% threshold)
(0.80 - avg(rate(node_cpu_seconds_total{mode!="idle"}[5m]))) / 0.80 * 100

# Memory headroom (bytes remaining before 90% usage)
avg(node_memory_MemAvailable_bytes) - (avg(node_memory_MemTotal_bytes) * 0.10)

# Request rate headroom (requests/sec before saturation)
10000 - sum(rate(http_requests_total[5m]))

# Time until saturation (weeks until CPU hits 80%)
(0.80 - avg(rate(node_cpu_seconds_total{mode!="idle"}[5m]))) /
  deriv(avg(rate(node_cpu_seconds_total{mode!="idle"}[5m]))[8w:]) /
  (7*24*3600)

Headroom summary report:

cat > capacity_headroom.md <<'EOF'
# Capacity Headroom Report (2025-02-09)

## Current Utilization
- **CPU**: 45% average (target: <80%)
- **Memory**: 62% (target: <90%)
- **Disk**: 71% (target: <95%)
- **Request Rate**: 4,200 req/s (capacity: 10,000)

## Headroom Analysis
- **CPU**: 35% headroom → ~12 weeks until saturation
- **Memory**: 28% headroom → ~16 weeks until saturation
- **Disk**: 24% headroom → ~8 weeks until full
- **Request Rate**: 5,800 req/s headroom → ~20 weeks until capacity

## Priority Actions
1. **Disk**: Implement log rotation or expand volume within 4 weeks
2. **CPU**: Plan horizontal scaling in next quarter
3. **Memory**: Monitor but no immediate action needed
EOF

→ Quantified headroom per resource w/ time-to-saturation.

If err: headroom already negative → reactive mode. Immediate scaling needed.

Step 4: Model Growth Scenarios

Factor in business projections:

# Example Python script for scenario modeling
import pandas as pd
import numpy as np

# Load historical data
df = pd.read_json('cpu_8weeks.json')

# Calculate weekly growth rate
growth_rate_weekly = df['value'].pct_change(periods=7).mean()

# Scenario 1: Current trend
weeks_ahead = 12
current_trend = df['value'].iloc[-1] * (1 + growth_rate_weekly) ** weeks_ahead

# Scenario 2: 2x user growth (marketing campaign)
accelerated_trend = df['value'].iloc[-1] * (1 + growth_rate_weekly * 2) ** weeks_ahead

# Scenario 3: New feature launch (+30% baseline)
feature_launch = (df['value'].iloc[-1] * 1.30) * (1 + growth_rate_weekly) ** weeks_ahead

print(f"Current Trend (12 weeks): {current_trend:.1%} CPU")
print(f"2x Growth Scenario: {accelerated_trend:.1%} CPU")
print(f"Feature Launch Scenario: {feature_launch:.1%} CPU")
print(f"Threshold: 80%")

→ Multiple scenarios → impact of business changes on capacity.

If err: scenarios exceed capacity → prioritize scaling before event.

Step 5: Generate Scaling Recommendations

Actionable recs:

## Capacity Scaling Plan

### Immediate Actions (Next 4 Weeks)
1. **Disk Expansion** [Priority: HIGH]
   - Current: 500GB, 71% used
   - Projected full date: 2025-04-01 (8 weeks)
   - Action: Expand to 1TB by 2025-03-15
   - Cost: $50/month additional
   - Justification: 5 weeks lead time needed

2. **Log Rotation Policy** [Priority: MEDIUM]
   - Current: Logs retained 90 days
   - Action: Reduce to 30 days, archive to S3
   - Savings: ~150GB disk space
   - Cost: $5/month S3 storage

### Near-Term Actions (Next Quarter)
3. **Horizontal Scaling - API Tier** [Priority: MEDIUM]
   - Current: 4 instances, 45% CPU
   - Projected: 65% CPU by 2025-05-01
   - Action: Add 2 instances (to 6 total)
   - Cost: $400/month
   - Trigger: When CPU avg exceeds 60% for 7 days

4. **Database Connection Pool** [Priority: LOW]
   - Current: 50 max connections, 40% used
   - Projected: 55% by Q3
   - Action: Increase to 75 in Q2
   - Cost: None (configuration change)

### Long-Term Planning (Next 6 Months)
5. **Migration to Auto-Scaling** [Priority: MEDIUM]
   - Current: Manual scaling
   - Action: Implement Kubernetes HPA (Horizontal Pod Autoscaler)
   - Timeline: Q3 2025
   - Benefit: Automatic response to load spikes

→ Prioritized list w/ costs, timelines, trigger conditions.

If err: recs rejected on cost → revisit thresholds or accept risk.

Step 6: Set Up Capacity Alerts

Alerts on low headroom:

# capacity_alerts.yml
groups:
  - name: capacity
    interval: 1h
    rules:
      - alert: CPUCapacityLow
        expr: |
          (0.80 - avg(rate(node_cpu_seconds_total{mode!="idle"}[5m]))) / 0.80 < 0.20
        for: 24h
        labels:
          severity: warning
        annotations:
          summary: "CPU headroom below 20%"
          description: "Current CPU headroom: {{ $value | humanizePercentage }}. Scaling needed within 4 weeks."

      - alert: DiskFillForecast
        expr: |
          predict_linear(avg(node_filesystem_free_bytes)[8w:], 4*7*24*3600) < 0.10 * avg(node_filesystem_size_bytes)
        for: 1h
        labels:
          severity: warning
        annotations:
          summary: "Disk projected to fill within 4 weeks"
          description: "Expand disk volume soon."

      - alert: MemoryCapacityLow
        expr: |
          avg(node_memory_MemAvailable_bytes) < 0.15 * avg(node_memory_MemTotal_bytes)
        for: 6h
        labels:
          severity: warning
        annotations:
          summary: "Memory headroom below 15%"

→ Alerts fire before saturation → time to scale proactively.

If err: alerts fire too often (alert fatigue) or too late (reactive) → tune thresholds.

Check

• Historical metrics ≥ 8 weeks
• predict_linear() returns sensible forecasts (no negatives)
• Headroom calc'd for all critical resources
• Scenarios include business projections
• Scaling recs have costs + timelines
• Capacity alerts configured + tested
• Report reviewed w/ eng leadership + finance

Traps

• Insufficient history: Linear predictions need 4+ weeks. Less → unreliable.
• Ignore step changes: Deploys, migrations, launches → spikes distort trends. Filter or annotate.
• Linear assumption: Not all growth linear. Exponential (viral) needs other models.
• Forget lead time: Cloud provision fast, but procurement, budgets, migrations take weeks. Plan early.
• No budget alignment: Planning w/o budget buy-in → last-min scramble. Involve finance early.

→

• setup-prometheus-monitoring — collect metrics used for capacity planning
• build-grafana-dashboards — viz forecasts + headroom
• optimize-cloud-costs — balance capacity vs cost