Load Balancing — L4 vs L7, Algorithms
Load balancing ensures traffic is distributed across multiple servers, improving availability, scalability, and reliability.
In system design, you’ll often mention load balancers (LBs) early when discussing horizontal scaling.
1. Why Load Balancing?
- Prevents a single server from being overwhelmed.
- Improves throughput and reduces latency.
- Enables fault tolerance (if one server dies, traffic is rerouted).
- Provides flexibility (rolling upgrades, canary deployments).
2. L4 vs L7 Load Balancing
Layer 4 (Transport Layer)
- Works at TCP/UDP level.
- Forwards packets without inspecting content.
- Very fast, low overhead.
Examples: AWS NLB, HAProxy (TCP mode).
Use cases: raw TCP/UDP traffic, gaming servers, VoIP.
Layer 7 (Application Layer)
- Works at HTTP/HTTPS level.
- Can inspect request content (headers, paths, cookies).
- Allows smart routing (e.g.,
/api→ API servers,/static→ CDN).
Examples: AWS ALB, Nginx, Envoy, Traefik.
Use cases: web applications, APIs, microservices.
3. Load Balancing Algorithms
3.1 Round Robin
- Cycles through servers sequentially.
- Simple and fair if servers are similar.
- Weakness: doesn’t account for server load differences.
3.2 Least Connections
- Routes to server with the fewest active connections.
- Good for variable request lengths.
- Can overload fastest servers if traffic is uneven.
3.3 Weighted Round Robin
- Assigns weights to servers (e.g., Server A = 2× capacity).
- Requests distributed proportionally.
3.4 Consistent Hashing
- Routes based on hash of client (e.g.,
userID % N). - Ensures same client/request goes to same server.
- Useful for caching & session stickiness.
- Weakness: requires rebalancing when servers change.
3.5 Random
- Randomly selects a backend.
- Surprisingly effective in some scenarios.
4. Advanced Load Balancing Features
- Health checks: detect and remove unhealthy servers.
- Sticky sessions: keep user bound to one server (not ideal for scaling).
- SSL termination: decrypt TLS at LB, offload CPU from servers.
- Rate limiting: throttle abusive clients.
- Global load balancing: route to nearest region (DNS-based, Anycast).
5. Trade-offs & Interview Tips
- L4 vs L7:
- L4 is faster, less flexible.
- L7 is smarter, more expensive.
- Algorithm choice:
- Round robin → uniform traffic.
- Least connections → variable workloads.
- Consistent hashing → caching/session workloads.
- Always mention health checks + failover.
- For global scale: mention CDNs + DNS load balancing.
6. Diagram (Conceptual)
Client Requests
|
+-----+-----+
| Load Balancer |
+-----+-----+
| |
Server A Server B7. Next Steps
- Explore CDNs & Edge Computing.
- Learn about Event-driven Architectures.