CDN Network - Scale and Constraints

Traffic Volume

Global Request Rate

• Aggregate requests: 10-50 million requests/second globally (equivalent to 10-15 trillion requests/month)
• HTTP GET dominance: 95%+ of requests are GET; remaining are HEAD, OPTIONS (CORS preflight), and POST (for edge functions)
• Request size distribution: Average request header ~800 bytes; average response varies by content type
• Connection reuse: 70-80% of requests served over persistent HTTP/2 or HTTP/3 connections

Bandwidth Per Edge Node

• Tier 1 PoPs (major metros: NYC, London, Tokyo): 1-10 Tbps aggregate capacity
• Tier 2 PoPs (secondary cities): 100 Gbps - 1 Tbps
• Tier 3 PoPs (smaller markets, ISP-embedded): 10-100 Gbps
• Per-server throughput: 40-100 Gbps per commodity server with DPDK/kernel bypass networking
• Typical utilization target: 40-60% of peak capacity to absorb traffic spikes

Total Egress

• Sustained global egress: 100-500 Tbps during peak hours
• Monthly data transfer: 5-20 exabytes/month for a top-tier CDN
• Egress cost structure: $0.01-0.08/GB depending on region (cheapest in NA/EU, most expensive in APAC/LATAM)

Storage Requirements

Cached Content Per PoP

• Hot tier (SSD): Top 5-10% of content by access frequency; serves 80-90% of requests
• Warm tier (HDD): Long-tail content; serves 10-20% of requests
• Memory cache: 256-512 GB RAM per server for ultra-hot objects (top 0.1%)

Origin Storage

• Total origin capacity: Varies by customer; CDN doesn't own origin content
• Origin shield cache: 50-200 TB per shield region (aggregates misses before hitting origin)
• Typical origin data set: 1 TB - 100 TB per large customer distribution

Metadata Storage

• Cache index per server: 50-100 GB in RAM (URL hash → disk location mapping)
• Per-object metadata: ~200 bytes (URL hash, content hash, TTL, headers, size, timestamps)
• Global configuration database: 10-50 TB (routing rules, certificates, customer configs)
• DNS zone data: 5-10 GB (millions of CNAME records, routing policies)

Network Bandwidth

Inter-PoP Communication

• Backbone links: 100 Gbps - 400 Gbps between major PoPs
• Purpose: Cache fill from parent/peer PoPs, configuration propagation, health data
• Traffic volume: 5-15% of total egress (most content served from local cache)
• Latency budget: <50ms between regional PoPs; <150ms cross-continent

Origin-Pull Bandwidth

• Cache miss rate: 5-15% of requests result in origin fetch
• Origin pull bandwidth: 5-50 Tbps globally (aggregated across all origins)
• Connection pooling: Maintain persistent connections to origins; 100-1000 connections per origin per PoP
• Origin shield reduction: Reduces origin load by 60-80% by consolidating cache fills

Client-Facing Bandwidth

• Last-mile delivery: Dominated by video (70-80% of bytes), images (10-15%), other (5-10%)
• Peak-to-average ratio: 3-5x (evening peak vs. overnight trough)
• Per-user bandwidth: 5-50 Mbps for video streaming; 1-5 Mbps for web browsing
• Protocol distribution: HTTP/2 (60%), HTTP/3 QUIC (25%), HTTP/1.1 (15%)

Geographic Distribution

Points of Presence (PoPs)

• Total PoPs: 200-400 globally (top CDNs like Cloudflare: 300+, Akamai: 4000+ edge servers)
• Tier 1 PoPs: 20-30 (major internet exchange points)
• Tier 2 PoPs: 50-100 (regional hubs)
• Tier 3 PoPs: 100-300 (last-mile, ISP-embedded)

Regional Distribution

Edge Location Strategy

• IXP co-location: Place servers at Internet Exchange Points for peering with multiple ISPs
• ISP-embedded: Deploy cache servers inside ISP networks (like Netflix Open Connect)
• Cloud region adjacency: PoPs near major cloud regions for fast origin pulls
• Population density: Correlate PoP placement with internet user density

Content Types and Sizes

Static Assets

Video Streaming

API Responses

• JSON API responses: 1-50 KB average; cacheable for 1-60 seconds
• GraphQL responses: 5-100 KB; often uncacheable due to personalization
• Webhook/event data: Typically not cached; passed through edge

Cache Hit Ratios and Origin Load Impact

Target Hit Ratios by Content Type

Impact on Origin Load

• Without CDN: Origin handles 50M req/s directly → requires massive origin infrastructure
• With 90% hit ratio: Origin handles 5M req/s → 10x reduction in origin capacity needed
• With 95% hit ratio: Origin handles 2.5M req/s → 20x reduction
• Bandwidth savings: 90% hit ratio saves $10-50M/month in origin egress costs

Factors Affecting Hit Ratio

• Content diversity: More unique URLs → lower hit ratio per object
• TTL configuration: Longer TTLs → higher hit ratio but staleness risk
• Query string handling: Ignoring irrelevant query params improves hit ratio
• Geographic spread: Popular content in one region may be cold in another
• Cache capacity: Larger caches hold more long-tail content

Cost Estimation

Bandwidth Costs (Monthly)

Storage Costs (Monthly)

Compute at Edge (Monthly)

Total Cost of Ownership

• Monthly operational cost: $50-80M for a global CDN
• Cost per GB delivered: $0.005-0.02 (internal cost)
• Customer pricing: $0.02-0.15/GB (margin depends on volume and region)
• Cost per million requests: $0.50-2.00

Peak Traffic Patterns

Live Events

• Major sporting events (Super Bowl, World Cup): 10-50x normal traffic spike
• Product launches (Apple keynote, game releases): 5-20x spike in specific regions
• Breaking news: 3-10x spike, geographically concentrated initially then spreading
• Ramp-up time: Seconds to minutes; must absorb without pre-warming

Flash Crowds

• Viral content: Single object goes from 0 to millions of requests/second
• Social media amplification: Tweet/post drives traffic to single URL
• Thundering herd on cache expiry: Many simultaneous requests for expired popular object
• Mitigation: Request coalescing, stale-while-revalidate, adaptive TTLs

Diurnal Patterns

• Peak hours: 7-11 PM local time (video streaming dominates)
• Business hours: 9 AM - 5 PM (web/API traffic peaks)
• Global smoothing: Follow-the-sun pattern means global peak is 2-3x global trough
• Weekend vs weekday: Video traffic 20-40% higher on weekends

Capacity Planning for Peaks

• Provisioning target: Handle 3-5x average traffic without degradation
• Burst capacity: Additional 2-3x via traffic shedding and graceful degradation
• Auto-scaling: Add capacity within minutes for sustained spikes
• Pre-positioning: Warm caches before known events (scheduled launches, sports)

Latency Budgets

End-to-End Latency Breakdown (Cache Hit)

End-to-End Latency Breakdown (Cache Miss)

SLA Targets

• Availability: 99.99% (52 minutes downtime/year)
• P50 latency (hit): <30ms
• P95 latency (hit): <75ms
• P99 latency (hit): <150ms
• Cache hit ratio: >90% for static content
• Purge propagation: <5 seconds globally