Load Balancing

A load balancer accepts client traffic and distributes it across a set of backend servers, removing unhealthy ones automatically. It gives you a single stable address in front of a changing pool of instances.

The externally-facing HTTP load balancers are implemented at the edge of Google's network rather than as a machine you run. A user connects to the nearest edge location and their request travels the rest of the way over the provider's private backbone, which is why a single global address can serve users worldwide with good latency.

Load balancing is where availability, latency and scaling all become concrete. It is also the topic most likely to appear as a selection question, because there are several products that look similar and only one is correct for a given scenario.

Being able to eliminate three of four options in a few seconds is worth more preparation time than almost any other networking topic, particularly on the Network Engineer and Architect exams.

1. 1Is the traffic coming from the internet or from inside your network? External or internal.
2. 2Do users need one address worldwide, or is one region enough? Global or regional.
3. 3Is it HTTP(S), or another protocol such as raw TCP or UDP? Application layer or network layer.
4. 4Does the load balancer need to terminate and inspect the connection, or pass it through unchanged? Proxy or pass-through.

Answer those four and the correct product is almost always determined. Most scenarios state at least two of them explicitly and imply a third.

• Backend service — defines the group of backends, the balancing mode, timeouts and the health check.
• Health check — probes backends and removes failing ones from rotation. A misconfigured health check that fails everything takes the whole service down, which is a common real-world outage.
• Backend group — usually a managed instance group or a network endpoint group, which is how serverless and container backends are attached.
• Forwarding rule — binds an IP address and port to the load balancer.
• URL map — routes requests to different backends by hostname and path, so one address can serve several services.
• Session affinity — sends a given client consistently to the same backend. Useful for stateful applications; it also undermines even distribution, so use it only when required.
• Capacity and balancing mode — controls when a backend is considered full and traffic should overflow to another region.

• A global web application on one anycast address with automatic failover between regions.
• One address serving several backend services routed by URL path.
• Internal load balancing in front of a service tier so callers use a stable private address.
• Gaming or telemetry servers over UDP where source addresses must be preserved.
• Blue/green and canary rollouts by weighting traffic between backend groups.