Edge Computing: Architecture Patterns and Trade-offs

Edge computing moves computation closer to the data source or end user. It is not a replacement for the cloud but a complement -- the right architecture places workloads where they deliver the most value based on latency, bandwidth, and data sovereignty requirements.

Edge Computing Tiers

Tier	Location	Latency	Examples
Device edge	On the device itself	< 1ms	Smartphones, sensors, gateways
Near edge	Local network / on-premises	1-10ms	Factory floor servers, retail stores
Far edge / MEC	Telco edge, regional PoP	5-20ms	5G MEC, ISP edge nodes
CDN edge	Distributed global PoPs	10-50ms	Cloudflare, CloudFront, Fastly
Cloud region	Centralized data center	50-200ms	AWS, GCP, Azure regions

Architecture Patterns

CDN Edge Computing

Run code at CDN points of presence, closest to end users.

Use cases: A/B testing, personalization, auth token validation, geo-routing
Platforms: Cloudflare Workers, Lambda@Edge, Vercel Edge Functions, Deno Deploy
Constraints: limited execution time, no persistent state, restricted APIs

IoT Edge

Process data on-premises or on gateway devices before sending to cloud.

Use cases: manufacturing quality inspection, predictive maintenance, video analytics
Platforms: AWS IoT Greengrass, Azure IoT Edge, Google Distributed Cloud Edge
Constraints: limited compute, unreliable connectivity, device management complexity

Multi-Access Edge Computing (MEC)

Compute at the telecom network edge, enabled by 5G.

Use cases: autonomous vehicles, AR/VR, real-time gaming, industrial automation
Platforms: AWS Wavelength, Google Distributed Cloud, Azure Edge Zones
Constraints: carrier partnerships required, limited geographic coverage

Edge vs Cloud Trade-offs

Dimension	Edge	Cloud
Latency	Very low	Higher (network round trip)
Bandwidth cost	Reduced (process locally)	Higher (transfer all data)
Compute capacity	Limited	Virtually unlimited
Data freshness	Real-time	Near-real-time to batch
Management complexity	High (many distributed nodes)	Lower (centralized)
Cost model	Hardware + maintenance	Pay-per-use
Security perimeter	Physically distributed	Centralized controls
Update deployment	Complex (fleet management)	Simple (centralized)

Edge AI/ML Inference

Running ML models at the edge enables real-time decisions without cloud round trips:

Model optimization -- TensorFlow Lite, ONNX Runtime, TensorRT for constrained devices
Hardware acceleration -- NVIDIA Jetson, Google Coral, Intel Neural Compute Stick
Model management -- versioning, A/B testing, and rollback across thousands of devices
Federated learning -- train on edge data without sending it to the cloud

When to Infer at the Edge

Factor	Edge Inference	Cloud Inference
Latency requirement	< 100ms	Seconds acceptable
Data sensitivity	Cannot leave premises	Can be sent to cloud
Connectivity	Intermittent or absent	Reliable
Model complexity	Simple to medium	Complex, large models
Update frequency	Infrequent	Frequent iteration

Data Synchronization Challenges

Edge architectures must handle data that exists in multiple locations:

Conflict resolution -- what happens when edge and cloud disagree? Last-write-wins, CRDTs, or manual resolution
Eventual consistency -- edge nodes may be offline; design for eventual sync
Data filtering -- send only aggregated or anomalous data to cloud, not raw streams
Bandwidth management -- prioritize critical data when connectivity is limited
State management -- where is the source of truth? Cloud, edge, or both?

CDN Edge Platforms Comparison

Platform	Runtime	Execution Limit	Global PoPs	Persistent Storage
Cloudflare Workers	V8 isolates	30s (paid)	300+	KV, R2, D1, Durable Objects
Lambda@Edge	Node.js, Python	30s (viewer), 30s (origin)	CloudFront PoPs	Limited (via S3/DynamoDB)
Vercel Edge Functions	V8 isolates	30s	Vercel network	Via external stores
Fly.io	Full containers	No limit	30+ regions	Volumes, LiteFS
Deno Deploy	V8 isolates	50ms CPU	35+ regions	Deno KV