While vertical scaling (bigger server) might work short-term, horizontal scaling with load balancing is the sustainable solution. It provides redundancy, allows incremental scaling, and doesn't require downtime.

Weighted Round Robin lets you assign Server A double the weight of B and C, so it receives proportionally more requests (2:1:1 ratio), matching its higher capacity.

With in-memory sessions and no session affinity, User A might log in on Server 1 but their next request goes to Server 2, which doesn't have their session. Use IP hash, sticky sessions, or externalized session storage.

With a 5-second interval and typical unhealthy threshold of 2-3 failures, it takes 10-15 seconds to mark the server as unhealthy. During this time, some requests fail. This is the trade-off between detection speed and avoiding false positives.

L7 load balancers operate at the HTTP layer and can inspect URL paths. L4 load balancers only see IP addresses and ports - they can't read HTTP content to determine the URL path.

Consistent hashing's key benefit: adding a server only moves ~1/N of the data (where N is the new server count). Adding server 6 moves ~17% of keys. With naive modulo hashing, ~83% would need to move!

A deep health check verifies not just that the application is running, but that critical dependencies (database, cache, external APIs) are also reachable. The /health endpoint should test these connections.

Gaming servers need minimal latency and maximum packet throughput. L4 load balancers are ideal - they simply forward packets without parsing HTTP content, adding minimal overhead.

Connection draining stops new requests to the server but allows existing connections to complete (up to a timeout). This prevents dropped connections and provides graceful shutdown for maintenance.

L7 load balancers can inspect HTTP headers, including Authorization tokens. You can decode the JWT, determine if it's a premium user, and route to the premium server pool.

Client-side load balancing eliminates the extra network hop through a load balancer and removes a potential single point of failure. The client gets the service registry and chooses a backend directly.

'Connection refused' means the port isn't listening. Either the health check is configured for the wrong port, or the application isn't binding to the expected port. A timeout error would indicate a different issue.

Round Robin ignores actual load. Server B takes 5x longer, so it accumulates 5x more concurrent connections. Use Least Connections instead - it adapts to actual server performance.

SSL termination means the LB decrypts traffic. If backends receive plain HTTP, traffic inside your network is unencrypted. For sensitive data, consider re-encrypting (SSL to backends) or using L4 passthrough.

With 5 servers: 10,000 / 5 = 2,000 req/s each. With 4 servers: 10,000 / 4 = 2,500 req/s each. That's a 25% increase per server. Ensure your servers have capacity headroom for failures!