Interactive scalable VPC topology and load balancing

Inventory Regions, AZs, tiers, peak ENIs and IPs, container or Lambda scaling, hybrid and peer CIDRs, protocols, static-IP needs, HTTP rules, inspection, failure scope, and growth horizon. Reserve nonoverlapping CIDR space and calculate usable addresses after AWS reservations and infrastructure consumption. A subnet name does not make it public; its associated route and resource addressing create the path. A larger CIDR added later cannot repair overlap with on-premises or every connected VPC.

Place public front doors only where internet reachability is required. Put application and data tiers in private subnets across enough AZs, and associate the intended route table with each subnet. Trace longest-prefix route at every routing domain and both forward and return controls. Plan route-table, target, and ENI growth before scale events. Centralization reduces management copies but can create throughput, route, cost, and failure concentration; distribute where the stated availability or path requirement decides.

ALB is Layer 7 for HTTP and HTTPS with host, path, and other HTTP-aware routing. NLB is Layer 4 for TCP, UDP, TLS, static-IP-oriented entry, and transport requirements. Gateway Load Balancer inserts and scales network virtual appliances transparently; it is not ordinary application ingress. Choose from protocol and function rather than a claim that one is universally faster.

Place targets across AZs and ensure target security controls, listener, health checks, ports, and routes align. Load balancer capacity cannot rescue exhausted target subnets or a database bottleneck. Health evaluation must keep routing state separate from application delivery. ALB documentation describes fail-open routing when every registered target is unhealthy: it sends requests to those targets despite health. An empty ALB target group instead has no registered target to receive the request; do not infer the all-unhealthy behavior from an empty registration set.

NLB documentation has two explicit fail-open conditions: all targets are unhealthy in all enabled Availability Zones, and the target group is empty. The empty-group condition still supplies no target and therefore no application endpoint; “fail open” describes the NLB routing mode, not a successful business response. Fixtures must record registration count, per-AZ health, routing decision, connection result, and application result separately. Capacity planning also includes target connections, long-lived flow distribution, draining, and enough surviving target capacity after an AZ failure.

For two directly connected nonoverlapping VPCs, peering can be simple but is nontransitive and requires routes on both sides. Transit Gateway fits a growing hub-and-spoke topology where attachment and centralized route-table management reduce a peering mesh. Neither choice overrides security groups, NACLs, DNS, or destination authorization. A shared services design needs explicit propagation and isolation, not one flat route domain.

The visual should progressively reveal Region, AZs, CIDRs, subnets, routes, load balancer, targets, SGs, NACLs, and connectivity. Inject IP exhaustion, wrong route association, asymmetric return, and target failure. Evaluate first failed gate and repair minimally. Security is not color coding: every permit or deny needs protocol, port, source, and direction. Performance is measured end to end; cost follows valid topology. Leave address and route headroom and test an AZ loss before calling the network scalable.

Load-balancer health behavior must be tested with registration state visible. For ALB, every registered target being unhealthy can produce fail-open routing, whereas an empty ALB group has no target. For NLB, both all-unhealthy-across-enabled-AZs and empty-group states are documented fail-open conditions; the empty case still cannot deliver an application response because there is no endpoint. Build fixtures for one healthy target, all registered targets unhealthy, and zero registered targets, then display routing mode, connection outcome, and application response independently. Health endpoints should validate the minimum dependency needed to serve while avoiding sensitive details. Address growth has a parallel cost test: larger subnets are not billed for empty addresses, but NAT, inspection, cross-AZ paths, load-balancer capacity, and idle surviving targets can be. Keep security-group references narrow enough that scaling adds authorized targets without opening an entire VPC CIDR.