Content Delivery and Edge Networking Questions

Requirements & goals

• Serve catalog pages and product JSON fast during promotional spikes; keep reasonably fresh inventory/pricing; minimize origin load and cost; simple operational model.

High-level cache layers

• CDN edge (global POPs) — serve HTML fragments, product JSON, images, JS/CSS.
• Regional cache layer (VPC e.g., CloudFront origin shield / regional cache) — group edge misses to reduce origin thundering.
• Application cache (Redis/Memcached) — hot product data, aggregations, search snippets.
• Database / origin (primary source) — authoritative.

TTL strategy

• Images/assets: long TTL (24h+) + immutable cache busting.
• Product JSON (catalog): short-to-medium TTL (30s–5m) depending on volatility; price/stock: very short (5–30s) or not cached for critical flows.
• HTML: vary by page type — landing/promo pages cached longer with SSR-staleness mechanisms.
• Use Cache-Control + stale-while-revalidate to serve slightly stale content while refreshing.

Cache warming

• Pre-warm CDN and regional caches before events: generate requests for top-N SKUs, promo landing pages, and API endpoints (parallelized, rate-limited).
• Use origin-shielded bulk refresh APIs where supported (e.g., CloudFront invalidation or CDN API prefetch).
• Automate from deployment CI/CD triggered by promo schedule.

Invalidation approaches

• Soft invalidation: send purge/ban for specific keys (product:{id}) on price/stock change; use CDN/push invalidation for affected paths.
• Cache-busting: update asset URLs or versioned API endpoints for large changes.
• Event-driven invalidation: integrate inventory/price change events to publish invalidation messages to a worker that updates Redis and issues CDN purges selectively.
• For high churn: use short TTL + no purge for frequent updates to avoid mass invalidation cost.

Trade-offs

• Freshness vs cost: very short TTLs + frequent purges increase origin load & CDN costs; use selective invalidation for hot items to balance.
• Complexity vs performance: event-driven selective invalidation and origin shield reduce load but add operational complexity.
• Consistency vs latency: stale-while-revalidate improves latency but serves slightly stale data briefly — acceptable for non-critical fields.
• Pre-warming reduces cold-start latency but requires accurate traffic prediction and automation.

Operational notes

• Monitor cache hit ratios, origin request rates, and purge costs; throttle warming and purges; maintain runbooks for emergency invalidation.