Load Balancer Class Decision: ALB vs NLB vs GLB

Three classes of cloud load balancer. The decision rule by use case with concrete numbers.

ALB / Application LB

The load balancer class decision is one of the foundational architecture choices for any cloud-deployed service. AWS offers Application Load Balancer (ALB), Network Load Balancer (NLB), and Global Accelerator-style global load balancing. Each fits different workloads; the choice depends on the protocol, performance requirements, and operational pattern of the service.

What ALB provides:

• Layer 7.: ALB operates at the application layer. It understands HTTP and HTTPS; it can inspect headers, paths, hostnames. The Layer 7 awareness enables application-aware routing.
• HTTP/HTTPS routing.: ALB terminates TLS and routes based on application-layer attributes. Multiple services can share an ALB; routing rules direct traffic to the right backend based on path or hostname.
• Host-based routing.: Different hostnames can route to different target groups. api.example.com to one backend; admin.example.com to another. The single ALB handles many services; the cost and operational footprint scale.
• Path-based routing.: Different URL paths can route to different target groups. /api/* to the API service; /static/* to the static asset service. The granular routing fits microservice architectures.
• Best for web applications.: ALB is the default choice for HTTP-based services. Web applications, REST APIs, GraphQL endpoints. The Layer 7 features match the workload's needs.
• Higher cost, more features.: ALB costs more per hour than NLB; it offers more features (WAF integration, OIDC authentication, response mutations). The cost reflects the feature set.

ALB is the right choice for HTTP and HTTPS workloads where Layer 7 routing matters.

NLB / Network LB

NLB operates at Layer 4 (TCP/UDP). It does not inspect application payloads; it forwards connections based on network-level attributes. The simplicity produces higher throughput, lower latency, and a different fit than ALB.

• Layer 4.: NLB operates at the transport layer. It does not parse HTTP; it forwards TCP and UDP connections. The Layer 4 simplicity produces lower overhead.
• TCP/UDP.: NLB supports TCP, UDP, and TLS pass-through. It works for any protocol that runs over TCP or UDP, not just HTTP. Database protocols, message queues, custom binary protocols all work.
• Connection-level.: NLB makes routing decisions per connection. Once a connection is routed, all packets follow the same path. The connection affinity simplifies stateful protocols.
• Higher throughput.: NLB handles millions of requests per second per AZ. The throughput ceiling is much higher than ALB; high-volume workloads benefit.
• Lower latency.: NLB adds milliseconds, not tens of milliseconds, to request latency. Latency-sensitive workloads benefit; the difference matters at scale.
• Best for high-performance non-HTTP workloads.: Database connections, message queue traffic, custom protocol services. The Layer 4 simplicity fits these workloads better than ALB's Layer 7 features.

NLB is the right choice for non-HTTP workloads or HTTP workloads where extreme throughput and low latency are the priority.

Global LB

Global load balancing routes traffic across regions. AWS Global Accelerator and Route 53 with geo-routing are the AWS options. The pattern adds complexity and cost; the value is geographic distribution and cross-region failover.

• Multi-region routing.: Traffic is routed to the closest healthy region. Users in Europe go to the EU region; users in the US go to the US region. The geographic affinity reduces latency.
• Geo-aware DNS or anycast.: Route 53 geo-routing uses DNS responses tailored to the requester's location. Global Accelerator uses anycast IPs that route to the nearest edge location. Both achieve geographic routing through different mechanisms.
• Best for global applications.: Applications with users distributed across continents benefit. The latency reduction from regional routing is significant; the user experience improves.
• Most expensive.: Global load balancing has higher per-hour costs than ALB or NLB alone. The architecture also typically duplicates infrastructure across regions; the total cost is significantly higher than single-region.
• Most complex.: Multi-region operations are operationally complex. Cross-region replication, regional failover, consistent deployment across regions all become concerns. The complexity is the primary cost.

Load balancer class decision is a per-service question. Different services in the same architecture often use different classes. Nova AI Ops integrates with load balancer telemetry, surfaces traffic patterns and costs, and helps teams identify when their LB choice does not match the workload's actual characteristics.