Google Cloud Load Balancing distributes traffic to application backends using managed networking infrastructure rather than an appliance that the customer must provision and maintain. Its purpose extends beyond spreading requests across servers: it can provide a suitable entry point, use health information to select backends, and support routing and security features appropriate to the protocol. A load balancer cannot fix a failed database or an application that returns incorrect data. Architects should therefore begin with the complete workload and its recovery expectations.
The service family matters. Google Cloud has Application Load Balancers and Network Load Balancers, including proxy and passthrough network models. These options differ by client protocol, public or private access, supported geographic scope, backend type, and traffic handling. A global frontend does not automatically guarantee successful regional failover, and a Layer 7 feature is not necessarily available on a Layer 4 product. The most useful design is one that can be explained and tested under real failure conditions.
Understand Application and Network Load Balancer differences
Application Load Balancers handle HTTP or HTTPS at Layer 7. They can consider hostname, URL path, and supported request attributes when deciding which backend service should handle a request. A web application can direct API requests and static assets to different services through an appropriate URL map. This convenience requires disciplined routing, certificate management, authorization, and request-level monitoring. In a shared frontend, a careless change can affect several otherwise independent applications.
Proxy Network Load Balancers handle supported Layer 4 TCP traffic through a managed proxy, which terminates the client-side connection and creates a backend-side connection. Passthrough Network Load Balancers preserve packet addressing and do not terminate the connection in a proxy, making them appropriate for supported protocols such as UDP and ESP. Passthrough and proxy modes differ in backend requirements and observable traffic behavior; they are not interchangeable settings. Confirm exactly how the client and backend see each connection.
A workload needing HTTP routing usually points toward an Application Load Balancer, while a TCP service with proxy requirements suggests a proxy Network Load Balancer. A service that must preserve client source information or handle certain non-TCP protocols may need passthrough. The Google Cloud architecture decision process helps evaluate these choices against application state, operating skills, cost, compliance requirements, and the supported feature matrix. No single load-balancer family is the default best choice for every protocol.
Choose a public or private frontend and the right geographic scope
An external frontend serves allowed clients outside the private network; an internal frontend provides access through supported private networking paths. A public application may need an external HTTPS entry point while its internal application services use separate private load balancing. Internal addressing can reduce exposure but does not replace identity or application authorization. A client with network reachability should not receive permission to read protected data solely because it arrived through an internal load balancer.
Global external Application Load Balancers can use a global anycast frontend and multi-region backends; regional modes restrict backend placement and have different operational characteristics. Internal Application Load Balancers also have regional and cross-region choices. Google’s current documentation calls global external passthrough Network Load Balancing a Preview capability as of November 2026. Do not design an unconditional production dependency on a preview feature without confirming its status, terms, and deployment constraints.
Where data must remain in particular regions, globally flexible routing can introduce unwanted residency, transfer, or operational complexity. Google Cloud DNS resolution and routing is another dependency: clients must resolve the intended frontend before any health or backend routing can matter. Record which names, regions, and backends are acceptable, including during an outage. Test the complete client path instead of treating the global address and its DNS record as separate administrative details.
Connect forwarding rules, backend services, and endpoint groups
The forwarding rule defines an IP address, port and protocol relationship, and load-balancing scheme. Depending on the chosen product, traffic may reach a proxy or passthrough backend selection. Backend services hold references to eligible backend resources, supported health checks and balancing policy. Application Load Balancers can add URL maps that select backend services by host or path. A working frontend is therefore the result of several connected configuration objects, each of which can be the source of an unexpected route.
Backends may be managed instance groups, network endpoint groups, or supported service attachments. Google Cloud virtual-machine backend choices matter when managed instance groups provide the serving capacity. GKE application endpoints introduce their own Kubernetes and pod lifecycle, while serverless network endpoint groups can connect supported Application Load Balancers to Cloud Run services. The backend service does not guarantee that its application has enough capacity or can reach its database. Each integration requires independent permissions and health criteria.
Draw one request path from DNS to the forwarding-rule address, target proxy or packet path, URL map when relevant, backend service, endpoint, and response. Mark where TLS terminates, which source addresses reach the backend, and which firewall rules permit health probes. That map distinguishes routing problems from application failures. It is particularly valuable when a deployment replaces backends or a regional outage changes which service instance should receive traffic.
Use health checks and autoscaling as evidence-based controls
A health check tests a configured condition, not the entire business transaction. A backend might return a simple success response while its database is unavailable or a critical downstream API is failing. Choose a probe that reflects meaningful readiness without creating unnecessary dependencies or high load. Health-check endpoints should be understood and monitored by both the application and network teams, since the probe’s result influences whether new traffic is likely to be directed to that backend.
Timeouts, intervals, and success or failure thresholds affect how quickly a backend changes status. Overly sensitive checks can eject capacity during brief load spikes; overly slow checks can send clients toward failure for too long. Google’s all-backends-unhealthy behavior varies by load-balancer mode. It is therefore inaccurate to promise that every type always removes all unhealthy destinations from consideration. Test one failing endpoint and then a complete backend outage in a safe environment to establish actual user impact.
A managed instance group may autoscale in response to load, but the load balancer does not instantly create unlimited healthy capacity. Scaling signals, backend startup time, quotas, and shared database limits determine whether growth helps. More web instances can worsen a constrained downstream database by multiplying connections. Record the minimum viable capacity and test how load balancing and autoscaling behave together, including what happens during an attempted failover or a badly timed application deployment.
Apply TLS, Cloud Armor, and CDN without confusing their roles
Application Load Balancers can use supported URL-map routing and TLS termination so one frontend can host multiple service paths. Teams need to define certificate ownership, how redirects and headers work, what traffic is allowed to backends, and where authorization happens. TLS can protect a connection while the application still grants inappropriate permissions. Likewise, correct backend routing does not establish that every API route is intended for every client. Treat proxy configuration and application security as complementary layers.
Google Cloud Armor protection can apply supported web-application firewall and related policies to suitable load balancer modes. It helps when rules are tied to a realistic threat model, tested for false positives, and monitored after activation. A WAF does not make an insecure application immune to attack or remove the need for identity checks and secure application code. Confirm product support for each Cloud Armor feature instead of assuming every frontend mode can run the same security policies.
Google Cloud CDN edge caching can reduce backend demand for eligible responses, but caching introduces correctness and privacy obligations. Cache keys, cookies, expiry, invalidation, and sensitive responses require deliberate configuration. A cache that treats personalized output as public can return information to the wrong client. Some identity-aware and CDN features have compatibility restrictions on the same backend service. Review those limits before enabling both, and verify behavior with authorized test users.
Treat session affinity and regional recovery as trade-offs
Session affinity can direct related requests toward a consistent backend in supported modes. This may help a legacy application that stores transient session state in memory, but affinity cannot guarantee that a session survives backend removal. It may also concentrate traffic unevenly. If the business requires continuity through instance replacement, persistent sessions and authoritative data generally need an appropriate shared or durable store. A sticky-session setting is not a backup or disaster-recovery plan.
Distribution methods and balancing modes are product-specific. Google Cloud does not promise simple round robin or least connections as universal behavior across all Application, proxy Network, and passthrough Network Load Balancers. Capacity, selected mode, health, proximity, and existing sessions can all affect observed traffic. Compare measured requests and latency across backends, then investigate any divergence from your design assumptions. An abstract claim that traffic is spread evenly is not enough to establish performance or fairness.
During a regional impairment, the surviving backend may lack the required data, secrets, identity services, or capacity. Global routing cannot correct these missing dependencies. Test a complete user transaction after a controlled backend or regional failure, not only the frontend status. Use the result to quantify actual recovery time, data consistency, and residual risk. Some applications are better served by a regional architecture plus tested disaster recovery than by unnecessary global complexity.
Monitor the complete request path and budget for it
Load-balancer monitoring can expose request counts, latency, backend health, and error rates. Correlate the frontend signals with application logs and Google Cloud Operations monitoring before declaring where a fault occurred. A burst of 5xx responses may be generated by an application backend; a timeout may come from networking, a quota, or a slow downstream dependency. Useful alerts identify the affected service, evidence, and appropriate owner rather than asking responders to interpret an unexplained graph.
Charges vary by load-balancer type and can include forwarding rules, regional traffic processed, proxy capacity for relevant internal modes, and separate internet transfer or service fees. Cloud Armor, CDN, Cloud NAT, and backend compute can add their own costs. Google separates global forwarding-rule pricing from regionally processed data charges, so a global address does not mean data flows without cost. Estimate the entire path using current regional pricing and realistic traffic volumes before choosing one configuration over another.
For an implementation review, state which protocols are required, which clients may reach the frontend, which backend regions are permitted, and what failure the design must withstand. Select the appropriate product and document forwarding rules, backend services, health checks, security, metrics, costs, and rollback. A controlled failure exercise should confirm whether a real user can finish the intended transaction. The objective is predictable service behavior during both normal traffic and abnormal conditions, not the highest possible number of optional features.