Interactive network flow, edge delivery, and transfer-cost tracing

Network optimization starts with a directional ledger. For each flow, record source and destination, protocol, public or private address, bytes in each direction, request count, Availability Zone and Region boundaries, account ownership, and every intermediary. A response can be much larger than its request, so collapsing directions hides the dominant charge. Preserve required multi-AZ resilience before attempting same-AZ placement. Then identify which hops perform processing—load balancer, NAT, Transit Gateway, endpoint, accelerator, or CDN—and which merely select an address. Apply compression, aggregation, caching, or co-location only after the exact path is visible. A route that avoids one processing service can introduce cross-AZ transfer elsewhere. The ledger should include failure paths as well as normal paths because traffic may shift Regions or traverse a different AZ during an outage. Use current service pricing for rates; the durable skill is multiplying each dated rate by the correct directional quantity without assuming that AWS-to-AWS traffic is free.

Route 53 answers DNS queries according to routing and health policy; it does not cache application payloads or provide a fixed application anycast address. CloudFront serves HTTP content through edge caches and can reduce origin requests and bytes when cache behavior is correct. Global Accelerator routes non-cacheable TCP or UDP traffic through static anycast entry to healthy endpoints and does not store objects. At the Regional front door, ALB provides HTTP and HTTPS Layer 7 routing, NLB provides Layer 4 transport including TCP and UDP with static-address use cases, and GWLB integrates virtual appliance fleets for inline inspection. These three ELB types are functional alternatives only when their protocol role matches. Current ELB billing includes load-balancer hours and type-specific capacity units, so compare traffic characteristics, processed bytes, connections or rules as applicable, and target fleet cost. ALB and NLB health behavior also affects routing: test registered-target health and application response rather than treating DNS reachability as success.

For cacheable downloads, estimate edge hit ratio by object and byte volume, not just request count. A high hit ratio can lower origin transfer and compute, while fragmented cache keys, short TTLs, or uncacheable personalization can erase that benefit. Include CloudFront requests, security or edge processing, invalidations, origin misses, and compression. For a dynamic non-HTTP control flow requiring stable entry addresses, Global Accelerator may justify its additional path charges; it loses if DNS steering alone meets failover and clients accept changing addresses. Health tests should exercise an unhealthy target, an empty target group, and an origin outage. A load balancer can route traffic while the application still fails, and documented fail-open conditions require degraded-path reasoning rather than a simplistic no-route answer. Track target health, response codes, connection errors, origin latency, cache status, and bytes at each boundary. The preferred design reduces total valid cost while preserving protocol, security, static-entry, latency, and resilience constraints. If the media later becomes private and uncacheable, direct origin or accelerator economics must be recalculated rather than inheriting the old CDN choice.