deangrant Load Balancing Algorithms
Use this skill when distributing traffic across backends.
1. Problem shape
Goals: availability, scale, latency. Homogeneous short HTTP vs variable work? Need affinity (sessions/cache) or fully stateless? Flash crowds may need admission before the LB.
2. L4 vs L7
- L7: content routing (URL, headers, cookies).
- L4: TLS passthrough, raw TCP/UDP, max PPS.
3. Algorithms (common)
| Algo | When |
|---|---|
| Round-robin / WRR | Homogeneous backends |
| Least connections | Variable request cost |
| Consistent hash / Maglev | Affinity, cache locality, graceful churn |
| Random / power-of-two | Simple good-enough with peer metrics |
Prefer stateless backends over sticky sessions when possible. Health checks must match readiness (not just liveness).
4. Quick checklist
- L4 vs L7 chosen deliberately.
- Algorithm matches workload skew.
- Affinity justified or eliminated.
- Health checks = ready to serve.
- Imbalance metrics watched (p99 per backend).
See reference.md and examples.md.