Network Load Balancer vs Application Load Balancer: the difference

TL;DR
- If the traffic is HTTP or HTTPS, pick ALB. NLB only earns its keep when the protocol isn't HTTP, when a client hardcodes an IP that can't change, or when you're publishing a PrivateLink endpoint.
- The difference that actually pages you at 2am isn't routing features — it's failure modes. ALB's 60-second idle timeout turns slow endpoints into 504s, and NLB's default TCP health check will keep routing traffic to an app that answers every request with a 500.
- Hourly list price is near-identical ($0.0225/hour for both in us-east-1); the bill diverges on LCU vs NLCU math, and the load balancer is rarely the biggest line on an SMB account anyway.
The verdict: ALB, unless the protocol says otherwise
Application Load Balancer is the default answer for anything serving HTTP or HTTPS to browsers, mobile clients, or other services. Host and path routing, HTTP-status-aware health checks, per-target 5xx metrics, WAF attachment, and OIDC/Cognito auth at the listener are all bundled into the LCU price. Rebuilding that behind an NLB means running nginx or Envoy yourself and paying for the EC2 hours anyway.
Network Load Balancer is the correct answer in a much narrower band of cases: raw TCP or UDP, a client that pins your IP inside a firewall rule, TCP source-IP preservation for licensing or legacy allowlists, or an endpoint service consumed over PrivateLink. On a typical SMB account that's one workload in ten.
The migration direction matters too. Moving NLB → ALB is usually a DNS cutover with a 60-second TTL and a week of running both (about $16 of overlapping LB hours at list). Moving ALB → NLB means losing WAF, host routing, and every error-rate metric you've been alerting on.
The decision tree, in one table
Start at the protocol question and stop at the first row that matches.
| Requirement | Pick | Why |
|---|---|---|
| HTTP/HTTPS, path or host routing to multiple services | ALB | Rules are free to add; each rule eval counts toward LCU |
| AWS WAF, Cognito/OIDC auth, fixed-response redirects | ALB | No WAF attachment point on NLB listeners |
| gRPC or WebSockets over HTTP/2 | ALB | End-to-end HTTP/2 on an HTTPS listener |
| Lambda targets (serverless backend) | ALB | NLB has no Lambda target type |
| Raw TCP, UDP, MQTT, syslog, SMTP, game traffic | NLB | ALB only speaks HTTP/HTTPS/gRPC |
| Client hardcodes an IP, or you need an EIP per AZ | NLB | One static IP per AZ; ALB node IPs rotate |
| TCP source IP must reach the target unchanged | NLB | ALB gives you X-Forwarded-For instead |
| Endpoint service for another VPC or account | NLB | PrivateLink requires NLB |
| Client IP allowlist at the app layer | NLB, or ALB + Global Accelerator | Accelerator gives static anycast IPs in front of ALB |
Where the two really diverge in production
| Dimension | ALB | NLB |
|---|---|---|
| Layer | 7 (HTTP/HTTPS/gRPC/WebSocket) | 4 (TCP/UDP/TLS) |
| Health checks | HTTP path + status matcher, app-aware | TCP port open by default; HTTP/HTTPS optional |
| Source IP at target | ALB node IP + X-Forwarded-For |
Client IP preserved unchanged |
| Static IP | No | One EIP per AZ |
| Security groups | Yes | Yes, since 2023 (older guides say no) |
| WAF | Attachable | Not supported |
| Access logs | Full request line, status, timings, ~every 5 min | TLS/connection metadata only |
| CloudWatch |
Leave a Reply