Penguin-inspired self-organizing load balancer with adaptive thermal eviction.

---

Emperor Penguins survive Antarctic blizzards without any central coordinator — each bird follows one rule: if you're cold, push inward; if you're warm, drift outward. The huddle self-organizes.

HuddleCluster applies this directly to server scheduling. Servers that run hot rotate to an outer ring to cool down. Cooled servers rotate back in. No manual tuning. No fixed thresholds. The cluster finds its own equilibrium.

---

Install

pip install huddle-cluster

Optional extras: fastapi · redis · grpc · kubernetes

---

Single-instance

from huddle_cluster import create_cluster
import requests

cluster = create_cluster([
    ("web-1", "10.0.0.1", 8080),
    ("web-2", "10.0.0.2", 8080),
    ("web-3", "10.0.0.3", 8080),
])
cluster.start()

with cluster.get_server_context() as server:
    response = requests.get(f"http://{server.host}:{server.port}/api")

What the cluster reports at any point:

print(cluster.health_report())

{
  "inner_servers": ["web-1", "web-3"],
  "outer_servers": ["web-2"],
  "fairness_score": 0.94,
  "rotation_count": 12,
  "requests_per_sec": 847.3,
  "cluster_health": "healthy"
}

---

Multi-node cluster

Coordinate a fleet of hosts — each node runs its own HuddleCluster; the master tracks enrollment, heartbeats, and health.

# Start the coordinator
huddle-cluster master start --port 7070

# Enroll nodes on each host
huddle-cluster agent start --id web-01 --master http://master:7070 --port 8080

# Inspect from anywhere
huddle-cluster nodes list

NODE ID                ADDRESS                STATUS       HB       LAST SEEN
─────────────────────────────────────────────────────────────────────────────
web-01                 10.0.0.1:8080          alive        142      0.8s ago
web-02                 10.0.0.2:8080          alive        139      1.1s ago
web-03                 10.0.0.3:8080          dead         41       34.2s ago

Ask the scheduler which node to send the next workload to:

curl http://master:7070/v1/scheduler/next

{ "ok": true, "node": { "node_id": "web-01", "address": "10.0.0.1", "port": 8080 } }

Live topology and Prometheus metrics are built in:

http://master:7070/dashboard      → real-time cluster topology
http://master:7070/v1/metrics     → Prometheus scrape endpoint
http://master:7070/v1/docs        → interactive API explorer (Swagger UI)

---

How it works

Concept	What it means
Inner ring	Active servers handling traffic right now
Outer ring	Servers cooling down after a hot streak
Thermal score	EMA of relative latency anomaly, CPU, memory, error rate
Rotation	Overheated servers evict outward; cooled servers return inward
Relative anomaly	Compared to the cluster median — adapts to any baseline automatically

No server is permanently marked bad. Every server gets rest and returns.

---

Performance

Under server failure, P95 latency stays under 86 ms where NGINX round-robin reaches 5,027 ms — a 58× reduction. Full methodology and results in the research paper below.

---

Documentation

Single-instance guide	USAGE.md
Cluster system	docs/CLUSTER.md — MasterNode, Scheduler, RBAC, dashboard, API
API explorer	http://your-master:7070/v1/docs (live, once the master is running)
Research paper	docs/HuddleCluster.pdf · arXiv preprint

---

Roadmap

• Thermal eviction, relative anomaly scoring, adaptive thresholds — v1.x
• Redis backend, gRPC routing, Kubernetes discovery, Prometheus, webhooks — v1.4
• Cluster system: MasterNode, AgentNode, CLI — v2.0
• Auto recovery, RBAC, metrics, dashboard, OpenAPI + Swagger UI — v2.x
• Cluster Scheduler — thermal-fitness workload placement — v3.0
• Cluster Auto Scaler — load-signal scale recommendations — v3.1
• Rolling Updater — zero-downtime batch upgrades with health gate — v3.2
• Service Discovery — health-aware registry, metadata-driven, DNS responder — v3.3
• HA Master — simplified Raft leader election, state replication, write redirect — v3.4
• Multi-Region — cross-datacenter topology, region-aware scheduling — v3.5
• Cluster Circuit Breaker — error-rate-based automatic trip/reset, scheduler exclusion — v4.0
• Rate Limiter — per-node token bucket, burst protection, scheduler exclusion — v4.1
• Canary Deployment — weight-based traffic splitting, start/advance/promote/abort — v4.2
• Observability — structured JSON logging, distributed trace IDs — v4.3, Level 4 complete
• TLS/HTTPS + mTLS, threaded HTTP server — v4.4, Level 5 (Production Hardening) in progress
• State persistence — HA term/voted_for + node registry survive restarts — v4.5
• mTLS node identity — client cert CN recorded on join — v4.6
• HA failover staleness fix + documented Raft limitations — v4.7
• OTLP log export (Jaeger/Tempo/OTel Collector compatible) — v4.8
• Docker + Kubernetes deployment manifests, SIGTERM graceful shutdown fix — v4.9
• WAN-latency simulation benchmark (partial — see docs for scope) — v4.10, Level 5 complete (7/7 addressed; 2 items — Raft hardening and WAN validation — improved but honestly still partial, see docs/CLUSTER.md)

---

Citation

Bhuiya, R. (2025). HuddleCluster: A Penguin-Inspired Self-Organizing Load Balancer
with Adaptive Thermal Eviction. https://github.com/rahadbhuiya/HuddleCluster

Bhuiya, Rahad (2026). HuddleCluster. figshare. Journal contribution.
https://doi.org/10.6084/m9.figshare.32397180

Bhuiya, Rahad (2026). HuddleCluster. Zenodo. https://doi.org/10.5281/zenodo.20348019

---

Author: Rahad Bhuiya  ·  License: MIT