---
id: wiki-2026-0508-edge-computing
title: Edge Computing
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Edge AI, Fog Computing, Edge Inference, MEC]
duplicate_of: none
source_trust_level: A
confidence_score: 0.92
verification_status: applied
tags: [edge, iot, distributed-systems, latency, on-device]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Rust/C++/Python
  framework: K3s / Cloudflare Workers / ONNX Runtime / WasmEdge
---

# Edge Computing

## 매 한 줄
> **"매 edge computing 의 핵심: compute moves to data — latency + bandwidth + privacy"**. 매 2010s CDN edge → 매 2020s function-edge (Cloudflare Workers, Deno Deploy, AWS Lambda@Edge), 매 2022 edge AI inference, 매 2024 5G MEC. 매 2026 현재 phone (Apple Intelligence, Pixel Gemini Nano) → CDN edge (Workers AI) → micro-DC 의 3-tier edge stack 의 일반화.

## 매 핵심

### 매 edge spectrum
- **Device edge**: 매 phone, sensor, MCU, robot.
- **Near edge / On-prem**: 매 factory floor, retail store K8s.
- **Far edge / CDN**: 매 Cloudflare, Fastly, Akamai PoPs.
- **MEC (Mobile Edge Computing)**: 매 5G base station co-located.

### 매 drivers
- **Latency**: 매 <10ms — 매 cloud round-trip 의 불가능.
- **Bandwidth**: 매 video / sensor stream의 cloud 전송 cost.
- **Privacy / sovereignty**: 매 data 의 region / device 외 미반출.
- **Resilience**: 매 offline operation.
- **Cost**: 매 egress fee 회피.

### 매 응용
1. Real-time vision (autonomous driving, AR).
2. Voice assistants (Siri on-device wake-word).
3. Industrial control (PLC + edge AI).
4. Game streaming, low-latency RTC.
5. Edge AI inference (Llama 3.2 on phone, Whisper on Pi).

## 💻 패턴

### Cloudflare Workers (function edge)
```ts
export default {
  async fetch(req: Request, env: Env): Promise<Response> {
    const cache = caches.default;
    const cached = await cache.match(req);
    if (cached) return cached;
    const data = await env.DB.prepare("SELECT * FROM posts LIMIT 10").all();
    const res = new Response(JSON.stringify(data),
      { headers: { "cache-control": "max-age=60" } });
    await cache.put(req, res.clone());
    return res;
  }
};
```

### Workers AI (edge LLM inference)
```ts
export default {
  async fetch(req, env) {
    const { prompt } = await req.json();
    const out = await env.AI.run("@cf/meta/llama-3.2-3b-instruct",
      { prompt, max_tokens: 200 });
    return Response.json(out);
  }
};
```

### K3s (lightweight K8s for edge)
```bash
# Install on edge node
curl -sfL https://get.k3s.io | sh -s - --disable traefik --node-name edge-01
# Deploy edge inference service
kubectl apply -f - <<EOF
apiVersion: apps/v1
kind: Deployment
metadata: { name: edge-yolo }
spec:
  replicas: 1
  template:
    spec:
      nodeSelector: { tier: edge }
      containers:
      - name: yolo
        image: ghcr.io/me/yolo11n:latest
        resources: { limits: { cpu: "2", memory: 2Gi } }
EOF
```

### ONNX Runtime on edge (Raspberry Pi 5)
```python
import onnxruntime as ort
import numpy as np
sess = ort.InferenceSession("yolov11n.onnx",
    providers=["CPUExecutionProvider"])
out = sess.run(None, {"images": img.astype(np.float32)})
```

### WasmEdge (portable edge runtime)
```rust
// Compile Rust → wasm32-wasi
fn main() {
    let img = std::fs::read("input.jpg").unwrap();
    let result = run_inference(&img);
    println!("{:?}", result);
}
// Run: wasmedge app.wasm
```

### MQTT-bridged hierarchical edge
```python
# Edge gateway aggregates sensors, forwards summaries to cloud
import paho.mqtt.client as mqtt

def on_local(client, _, msg):
    summary = aggregate(msg.payload)
    cloud.publish("plant/summary", summary)

local = mqtt.Client(); local.connect("localhost"); local.on_message = on_local
local.subscribe("sensors/#"); local.loop_start()

cloud = mqtt.Client(); cloud.connect("cloud.broker.com")
```

### CRDT-based offline-first sync
```ts
import * as Y from "yjs";
import { WebsocketProvider } from "y-websocket";
const doc = new Y.Doc();
// Works offline, merges automatically when reconnected
new WebsocketProvider("wss://sync.edge.local", "doc1", doc);
```

### Edge LLM with quantization (mobile)
```python
# Convert Llama 3.2 3B to 4-bit GGUF for mobile
# llama.cpp or MLX
from mlx_lm import convert
convert("meta-llama/Llama-3.2-3B-Instruct",
        mlx_path="llama-3b-q4", quantize=True, q_bits=4)
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Static asset / API cache | CDN (Cloudflare, Fastly) |
| Low-latency API | Workers / Deno Deploy |
| Stateful edge service | K3s on near-edge |
| Sensor / IoT | MQTT + edge gateway |
| Edge AI inference | ONNX Runtime / TFLite / Core ML |
| Cross-platform portable | Wasm (WasmEdge, Spin) |
| Personal AI | On-device LLM (MLX, llama.cpp) |

**기본값**: 매 latency-sensitive read 은 CDN edge, 매 sensor data 의 edge gateway 에서 aggregate.

## 🔗 Graph
- 부모: [[Distributed Systems]]
- 변형: [[CDN]] · [[Serverless]] · [[MEC]]
- 응용: [[클라우드 인프라 및 IaC 운영 표준|IoT]] · [[Edge AI]] · [[On-device AI]]
- Adjacent: [[Cloud Native]] · [[Data Privacy & Local Processing]]

## 🤖 LLM 활용
**언제**: 매 edge deployment scaffolding (K3s manifests, Workers code), 매 quantization workflow, 매 latency-budget reasoning.
**언제 X**: 매 hard-real-time (<1ms) — 매 LLM 의지 X. 매 deterministic timing 의 RTOS 사용.

## ❌ 안티패턴
- **Edge as cloud-with-extra-steps**: 매 unnecessary 사용 — 매 latency / privacy 명확한 driver 없으면 cloud.
- **No degradation strategy**: 매 edge offline → 전체 fail.
- **State at every edge**: 매 consistency nightmare — 매 CRDT / eventual 의 사용.
- **Same model size as cloud**: 매 device OOM — 매 quantize + distill.
- **Ignoring network partitions**: 매 split-brain → corrupted state.
- **Pushing all logic to client**: 매 trust boundary violation.

## 🧪 검증 / 중복
- Verified (Cloudflare Workers docs 2025, K3s docs, ETSI MEC spec, ONNX Runtime mobile docs, Llama 3.2 release notes 2024).
- 신뢰도 A.
- 관련: [[Cloud Native]], [[Data Privacy & Local Processing]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — edge spectrum + 2026 tooling (Workers AI, MLX, K3s) |