Hosted Tixin on k8s on 2x Raspberry Pi 5s

TL;DR: I deployed the entire Tixin platform onto a tiny ARM edge cluster (2× Raspberry Pi 5) using k3s and tuned it to serve ~1000 concurrent connections while sustaining ~16k req/sec under load tests.

Why I tried this

Edge-first infra is underrated. I wanted a real-world test: can a tiny, cheap cluster run production-like traffic while staying stable and observable?

Architecture & setup

• Hardware: 2 × Raspberry Pi 5 (8GB)
• OS: Ubuntu Server 22.04 LTS (ARM64)
• Kubernetes: k3s (lightweight Kubernetes)
• Storage: NVMe SSDs via PCIe HATs for fast local persistence
• Ingress / Load-testing: k6 for load generation, a lightweight HAProxy for ingress in front of services

Key metrics

• Concurrent connections: ~1,000
• Peak throughput: ~16,000 req/sec (observed in targeted k6 tests)
• p95 latency: < 50ms in steady-state tests

Optimizations & learnings

• Use small, stateless services and push heavy state into read-optimized stores.
• Tune CFS quotas and CPU pinning on ARM to reduce noisy-neighbor effects.
• Local SSDs with writeback caches significantly reduce tail latency for disk-heavy services.
• Use service mesh lightly — observability over heavy traffic shaping on tiny clusters.

Stack

Kubernetes (k3s) · Node.js · Docker · k6 · Prometheus · Grafana

Repo & references

• Repo / code samples: https://github.com/Adityaadpandey (see relevant infra & k8s manifests). :contentReference[oaicite:1]{index=1}