55  Edge Architecture Review

In 60 Seconds

Edge computing architecture follows a three-tier model (edge, fog, cloud) where edge gateways perform Evaluate-Format-Reduce (EFR) functions at Level 3 of the Cisco IoT Reference Model. Architecture decisions depend on latency requirements (edge for <10ms), bandwidth constraints, privacy needs, and the distinction between Massive IoT (high volume, delay tolerant) and Critical IoT (low latency, high reliability).

55.1 Learning Objectives

Time: ~15 min | Level: Intermediate | Unit: P10.C10.U02

Key Concepts

  • Three-tier IoT architecture: The canonical edge/fog/cloud layering where device-level microcontrollers, gateway-level fog nodes, and cloud servers each handle processing tasks matched to their resource capabilities and latency requirements.
  • Partition decision: The choice of which pipeline stage marks the boundary between local (edge/fog) and remote (cloud) processing, optimised to meet latency, bandwidth, and accuracy requirements simultaneously.
  • Protocol bridging: The translation of device-side protocols (Modbus, BLE, Zigbee) to cloud-side protocols (MQTT, HTTPS) at the gateway tier, enabling heterogeneous devices to connect to standard cloud platforms.
  • Edge node taxonomy: Classification of edge processing hardware by resource tier: Class 1 (microcontroller, 32 KB–4 MB, <100 MHz), Class 2 (SBC/SoC, 64 MB–4 GB, 1–2 GHz), Class 3 (x86 gateway, 4–64 GB, 2–4 GHz).
  • Southbound/northbound interfaces: Southbound interfaces connect the gateway to field devices (sensors, actuators); northbound interfaces connect the gateway to cloud or enterprise systems; both must be explicitly designed and secured.

By the end of this chapter, you will be able to:

  • Identify Architecture Patterns: Distinguish between edge, fog, and cloud computing tiers and their roles
  • Apply Decision Frameworks: Use structured decision trees to select optimal processing locations
  • Evaluate Trade-offs: Assess latency, bandwidth, privacy, and cost trade-offs for different architectures
  • Match Use Cases: Map IoT application requirements to appropriate architectural patterns
  • Compare Massive vs Critical IoT: Distinguish the distinct requirements of high-volume delay-tolerant versus low-latency mission-critical deployments

55.2 Prerequisites

Required Reading:

Related Chapters:

Think of IoT architecture like choosing where to cook a meal:

  • Edge (your kitchen): Fast, private, but limited equipment
  • Fog (neighborhood restaurant): More equipment, serves multiple homes
  • Cloud (factory kitchen): Massive scale, can serve thousands, but delivery takes time

The right choice depends on what you are cooking (your data), how quickly you need it (latency), and how many people you are serving (scale). This chapter helps you make that choice systematically.


55.3 Core Concepts Summary

55.3.1 Edge Computing Terminology

Concept Definition Key Characteristics
Edge Computing Processing data near its source (IoT devices) Low latency (<10ms), reduced bandwidth, local autonomy
Fog Computing Intermediate layer between edge and cloud Aggregation, preprocessing, hierarchical architecture
Cloudlet Small-scale data center at network edge VM-based edge services, mobile edge computing
MEC (Multi-access Edge Computing) Telco edge infrastructure at cell towers 5G integration, ultra-low latency, operator-managed
Edge Gateway IoT Reference Model Level 3 Evaluate, format, reduce data before cloud
Data in Motion Streaming data from devices (Levels 1-3) Event-driven, real-time processing
Data at Rest Stored data in databases (Level 4+) Batch processing, historical analysis

55.3.2 IoT Reference Model: Edge-Relevant Levels (1-4)

The Cisco IoT Reference Model defines seven levels from physical devices (Level 1) through collaboration and processes (Level 7). For edge architecture decisions, the first four levels are most relevant because they determine where data is processed before reaching the application layers. See IoT Reference Models for the complete seven-level model.

Level Name Function Processing Type
Level 1 Physical Devices & Controllers Sensors, actuators, embedded systems Local sensing, basic control
Level 2 Connectivity Communication protocols, gateways Reliable data transport
Level 3 Edge Computing Data reduction, filtering, formatting Real-time preprocessing
Level 4 Data Accumulation Storage systems, databases Long-term storage, batch analytics

55.3.3 Edge Gateway Three Functions (EFR Model)

Function Purpose Examples
Evaluate Filter low-quality or invalid data Discard out-of-range values, check sensor health
Format Standardize data for cloud ingestion Convert units, normalize timestamps, add metadata
Reduce (Distill) Minimize data volume before transmission Downsample, aggregate, compute statistics

55.4 Edge vs Cloud Trade-off Matrix

55.4.1 Architectural Trade-offs

Factor Edge Cloud Hybrid
Latency <10ms (local) 50-200ms (network delay) Variable (critical=edge, batch=cloud)
Bandwidth Cost Low (minimal WAN traffic) High (continuous uploads) Medium (selective sync)
Processing Power Limited (constrained devices) Unlimited (elastic scaling) Flexible (offload when needed)
Data Privacy Local (stays on-premise) Concerns (third-party servers) Selective (sensitive=edge)
Maintenance Distributed (physical access) Centralized (remote updates) Complex (dual management)
Storage Capacity Limited (local disks) Petabyte-scale Tiered (hot=edge, cold=cloud)
Reliability Single point of failure Redundant infrastructure Failover capability
Cost Model CapEx (upfront hardware) OpEx (pay-as-you-go) Mixed

55.4.2 Use Case Suitability

Application Best Architecture Justification
Autonomous Vehicles Edge-heavy <10ms latency required for safety, no cloud dependency
Predictive Maintenance Hybrid Edge for anomaly detection, cloud for fleet-wide trends
Smart Home Edge-heavy Privacy, works during internet outages
Video Surveillance Hybrid Edge for motion detection, cloud for long-term storage
Agricultural Monitoring Hybrid Edge for irrigation control, cloud for historical analysis
Industrial Safety Edge-heavy Real-time shutdowns, no network dependency
Customer Analytics Cloud-heavy Aggregate across locations, complex ML models
Energy Optimization Hybrid Edge for load balancing, cloud for demand forecasting

55.5 Architecture Patterns

55.5.1 Hierarchical Edge-Fog-Cloud Architecture

Three-tier architecture diagram with IoT devices at the bottom connecting to edge gateways for local filtering, fog nodes in the middle for regional aggregation, and cloud services at the top for centralized analytics and storage
Figure 55.1: Three-tier edge-fog-cloud architecture showing data flow from devices through edge gateways, fog aggregation nodes, to cloud storage and analytics

55.5.2 Processing Location Decision Tree

Flowchart decision tree starting with latency requirement check, branching to edge processing for sub-10ms needs, fog processing for 10-100ms needs, and cloud processing for latency-tolerant workloads, with cost and capability trade-offs annotated at each branch
Figure 55.2: Decision tree for selecting processing location based on latency requirements, showing trade-offs between edge (fast, hardware costs), fog (balanced), and cloud (flexible, bandwidth costs) processing

55.5.3 Data Reduction Pipeline

Pipeline diagram showing raw sensor data entering the Evaluate stage where invalid readings are discarded, then the Format stage where data is normalized, and finally the Reduce stage where aggregation compresses data volume by orders of magnitude before cloud transmission
Figure 55.3: Edge gateway data reduction pipeline demonstrating 1000x volume reduction through evaluation, formatting, and aggregation stages


55.6 Massive IoT vs Critical IoT

The 3GPP standards body formally distinguishes two IoT traffic profiles that demand fundamentally different edge architectures. Getting this distinction wrong leads to either over-engineering (spending $50 per node on hardware that only needs $2 worth) or under-engineering (deploying delay-tolerant infrastructure for life-safety applications).

Characteristic Massive IoT Critical IoT
Scale 100,000+ devices per km2 10-1,000 devices per site
Latency tolerance Seconds to minutes Sub-10ms mandatory
Reliability Best-effort (99%) Ultra-reliable (99.999%)
Data rate per device <100 kbps 1-100 Mbps
Power Battery (5-10 year life) Mains-powered acceptable
Cost per node <$5 target $50-$500 acceptable
Example protocols NB-IoT, LoRaWAN, Sigfox 5G URLLC, TSN, PROFINET
Example applications Smart metering, agriculture, asset tracking Autonomous vehicles, surgical robots, industrial safety

Why this matters for architecture: A smart agriculture deployment with 50,000 soil sensors reporting hourly moisture readings (Massive IoT) can tolerate cloud-only processing because a 500ms delay is irrelevant when readings arrive every 3,600 seconds. In contrast, a robotic welding cell where a collision must trigger emergency stop within 1ms (Critical IoT) cannot involve any network hop beyond the local edge controller.

Real-world hybrid example: A smart factory floor typically runs both profiles simultaneously. Thousands of environmental sensors (temperature, humidity, vibration) operate as Massive IoT through LoRaWAN gateways with cloud analytics. Meanwhile, 20-30 robot safety interlocks operate as Critical IoT through local PLCs with sub-millisecond response. The two profiles share the same physical space but use completely different communication stacks, processing tiers, and reliability budgets.

Massive IoT vs Critical IoT: Quantifying the Architectural Differences

Consider a smart factory with both traffic types. The math shows why they need separate infrastructure:

Massive IoT Network (10,000 environmental sensors): \[ \text{Data Rate per Device} = \frac{100 \text{ bytes}}{3600 \text{ seconds}} = 0.028 \text{ bytes/sec} \] \[ \text{Total Aggregate Rate} = 10{,}000 \times 0.028 = 280 \text{ bytes/sec} = 24 \text{ MB/day} \] \[ \text{Latency Budget} = 60 \text{ seconds (acceptable delay)} \] \[ \text{Reliability Target} = 99\% \text{ (packet loss tolerable)} \]

Critical IoT Network (30 safety interlocks): \[ \text{Data Rate per Device} = \frac{50 \text{ bytes}}{0.001 \text{ seconds}} = 50{,}000 \text{ bytes/sec} \] \[ \text{Total Aggregate Rate} = 30 \times 50{,}000 = 1.5 \text{ MB/sec} = 130 \text{ GB/day} \] \[ \text{Latency Budget} = 1 \text{ ms (hard real-time)} \] \[ \text{Reliability Target} = 99.999\% \text{ (< 1 failure per 100,000)} \]

Cost per Device: \[ \text{Massive IoT} = \$5 \text{ hardware} + \$1/\text{year connectivity} = \$6 \text{ (amortized)} \] \[ \text{Critical IoT} = \$300 \text{ hardware} + \$0 \text{ connectivity (wired)} = \$300 \]

Network Capacity Requirements: \[ \text{Massive IoT LoRaWAN} = \frac{10{,}000 \text{ devices}}{1{,}000/\text{gateway}} = 10 \text{ gateways at } \$500 = \$5{,}000 \] \[ \text{Critical IoT TSN/PROFINET} = \frac{30 \text{ devices}}{30/\text{PLC}} = 1 \text{ PLC at } \$15{,}000 = \$15{,}000 \]

The 50x cost difference per device and 3x gateway cost difference justify separate architectures – mixing them would either over-provision Massive IoT ($50 to $300/node) or under-provision Critical IoT (cannot meet 1ms latency).

55.6.1 IoT Profile Cost Explorer

Use this calculator to compare infrastructure costs for Massive IoT versus Critical IoT deployments at different scales.

Side-by-side comparison diagram contrasting Massive IoT deployments with thousands of low-cost battery-powered sensors using LoRaWAN against Critical IoT deployments with fewer high-reliability wired devices using 5G URLLC or TSN, highlighting differences in latency, cost, and reliability requirements
Figure 55.4: Comparison between Massive IoT (high volume, delay tolerant) and Critical IoT (low latency, high reliability) deployment patterns

55.7 Workload Placement Decision Framework

55.7.1 Comprehensive Decision Tree

Multi-level decision tree guiding architecture selection through questions about latency requirements, data volume, privacy constraints, and cross-site analytics needs, with terminal nodes recommending edge-only, fog-assisted, hybrid, or cloud-centric architectures
Figure 55.5: Decision tree for selecting edge, cloud, or hybrid IoT architecture based on application requirements

55.7.2 Workload Placement Guidelines

Workload Type Best Location Justification
Safety-critical control Edge Cannot depend on network
Immediate alerting Edge <10ms response needed
Data filtering/validation Edge Reduce bandwidth 10-100x
Local dashboards Edge Works during outage
Sensor fusion Edge/Fog Combine multiple inputs locally
Anomaly detection (rules) Edge Simple thresholds
Anomaly detection (ML) Fog GPU for inference
Historical analysis Cloud Access to all locations
Model training Cloud Massive datasets, compute
Regulatory reporting Cloud Centralized compliance
Cross-site optimization Cloud Global view needed

55.8 Case Study: Walmart’s Edge-Fog-Cloud Architecture for Refrigeration Monitoring

Walmart operates over 4,700 US stores, each containing 50-100 refrigeration units that must maintain temperatures between -18C (frozen) and 4C (fresh). A single refrigeration failure can destroy $10,000-$35,000 in perishable inventory within hours. Before 2019, temperature monitoring relied on manual walk-through checks twice daily – a 12-hour gap where failures went undetected.

Architecture deployed (2019-2021):

Tier Components Processing Latency
Edge (per unit) Wireless temp sensor + BLE gateway Threshold alerting: triggers local alarm if temp exceeds limits for >5 minutes <500ms
Fog (per store) In-store edge server (Dell Edge Gateway 5100) Aggregates 50-100 sensors, runs compressor anomaly detection model (gradient boosted tree, 12 MB), correlates door-open events with temp spikes 2-5 seconds
Cloud (centralized) Azure IoT Hub + Databricks Fleet-wide pattern analysis across 4,700 stores, seasonal demand forecasting, predictive maintenance scheduling 30-60 seconds

Why not cloud-only? Walmart calculated that sending raw temperature readings (every 30 seconds from 350,000+ units) to the cloud would cost approximately $2.1 million per year in cellular data alone. The fog tier reduces data volume by 97% – only anomalies, hourly summaries, and compressor health scores are transmitted to the cloud.

Why not edge-only? The compressor anomaly detection model requires correlating temperature trends with door-open frequency, defrost cycles, and ambient store temperature – data from multiple sensors that only the fog tier can aggregate. The edge sensor has 64 KB RAM and cannot run the model.

Results after 18 months:

  • $86 million saved in prevented inventory loss (vs. manual-check baseline)
  • Compressor failures predicted 3-7 days early in 73% of cases (fog-tier model)
  • Store energy costs reduced 8% through cloud-tier optimization of defrost scheduling across climate zones
  • Mean time to detect refrigeration failure: reduced from 6 hours to 47 seconds

Architecture lesson: The three-tier split was driven by three distinct latency requirements – immediate local alerting (edge), real-time anomaly detection (fog), and fleet optimization (cloud). Each tier earned its place by solving a problem the other tiers could not.

Decision framework applied to this case:

Question Answer Architecture Implication
Is sub-second response required? Yes (compressor failure alert) Edge tier mandatory
Does processing need multi-sensor correlation? Yes (temp + door + defrost + ambient) Fog tier mandatory
Is fleet-wide optimization valuable? Yes (defrost scheduling across climate zones) Cloud tier mandatory
Can you tolerate internet outage? No (food safety regulations require continuous monitoring) Edge must operate independently
What is the data reduction ratio? 97% (only anomalies and summaries to cloud) Fog tier pays for itself in bandwidth savings

Common Pitfalls

Specifying an edge processing pipeline without measuring its memory footprint, CPU utilisation, and power draw on the actual target hardware will lead to components that fail to fit or drain the battery faster than expected.

A single binary handling sensor acquisition, protocol conversion, analytics, and cloud communication fails catastrophically when any one function encounters a bug or resource exhaustion. Structure edge software as independent, supervised processes.

Much attention goes to securing the southbound device connection, but northbound cloud credentials stored in plaintext on a gateway are an equally serious vulnerability. Use secure key storage for all northbound cloud authentication credentials.

Every edge architecture review must specify what happens when the network is unavailable, the cloud is unreachable, a sensor fails, or the gateway reboots mid-transmission. Without defined failure modes, production incidents become unrecoverable.

55.9 Summary and Key Takeaways

55.9.1 Core Architecture Principles

  1. Process data as close to the source as possible to minimize latency and bandwidth
  2. Edge gateways perform Evaluate-Format-Reduce (EFR) functions before cloud upload
  3. The Cisco IoT Reference Model has 7 levels (devices through collaboration); edge architecture decisions focus on Levels 1-4
  4. Massive IoT (high volume, delay tolerant) vs Critical IoT (low latency, high reliability) require different architectures
  5. Hybrid architectures combine edge real-time processing with cloud analytics

55.9.2 Architecture Selection Summary

Aspect Edge Advantage Cloud Advantage
When to use Real-time, safety-critical, privacy-sensitive Historical analysis, complex ML, cross-site aggregation
Data strategy Filter and reduce locally Store everything for long-term insights
Processing Simple rules, lightweight ML Deep learning, big data analytics
Failure mode Local autonomy during outage Redundant, always-available infrastructure

Key Takeaway

Edge computing architecture follows five core principles: process data close to the source, use Evaluate-Format-Reduce (EFR) at gateways, understand the 7-level Cisco IoT Reference Model (with Levels 1-4 driving edge decisions), distinguish Massive IoT from Critical IoT requirements, and default to hybrid architectures that combine edge real-time processing with cloud analytics. The right architecture depends on latency needs, bandwidth constraints, privacy requirements, and whether your deployment prioritizes scale or safety.

“Choosing the Right Kitchen!”

The Sensor Squad was planning a big meal for the entire Smart City, and they needed to figure out where to cook.

“I will cook everything in my tiny kitchen!” said Max the Microcontroller (the edge). “I am super fast – I can make a sandwich in 10 milliseconds!”

“But your kitchen is small,” said Sammy the Sensor. “You can only make simple meals. What about the fancy AI-powered recipes?”

“That is where I come in!” announced the Cloud, a massive factory kitchen in the sky. “I can cook for thousands of people and use the fanciest recipe books. But delivering food takes 200 milliseconds because I am far away.”

“And I am the neighborhood restaurant!” said the Fog Node. “I am bigger than Max’s kitchen but closer than the Cloud. I can handle medium-sized jobs.”

Lila the LED drew a diagram on the board: “So here is how we decide:”

  • “Need food in 10 milliseconds? (Emergency!) Cook at Max’s kitchen (EDGE)”
  • “Need to analyze recipes from 100 restaurants? Use the Cloud!”
  • “Need something in between? The Fog restaurant helps!”

Bella the Battery added the clincher: “And for safety – like detecting a gas leak – ALWAYS cook at the edge. You cannot wait for a delivery from the Cloud when there is an emergency!”

“That is called a HYBRID approach,” Max explained. “I handle the fast, important stuff. The Cloud handles the big, brainy stuff. Together we are unstoppable!”

What did the Squad learn? IoT architecture is like choosing where to cook: edge for speed and safety, cloud for scale and smarts, and hybrid for the best of both worlds!

55.10 Knowledge Check

A retail chain with 500 stores deploys 20 cameras per store (10,000 cameras total) for customer analytics. Each camera generates 1080p video at 30 FPS. Design a three-tier edge-fog-cloud architecture showing data flow and reduction at each tier.

Raw data generation (per camera):

  • 1080p @ 30 FPS: ~3 Mbps bitrate (H.264 compression)
  • Per store (20 cameras): 20 x 3 Mbps = 60 Mbps = 7.5 MB/second
  • Per store per day: 7.5 MB/s x 86,400 seconds = 648 GB/day
  • Total 500 stores: 648 GB x 500 = 324 TB/day raw video

Tier 1 - Edge (in-store processing on NVIDIA Jetson Nano):

  • Run person detection CNN (YOLOv5) at 30 FPS per camera
  • Extract metadata: person count, dwell time, heatmap coordinates (not full video frames)
  • Per camera: 30 FPS x 50 bytes metadata = 1,500 bytes/second
  • Per store: 20 cameras x 1,500 = 30,000 bytes/s = 30 KB/s
  • Reduction: 7.5 MB/s to 30 KB/s = 250x reduction at edge

Tier 2 - Fog (regional server covering 50 stores):

  • Aggregate foot traffic patterns across 50 stores in region
  • Compute hourly store visit statistics: peak hours, avg dwell time, section popularity
  • Per region uplink: 50 stores x 30 KB/s = 1.5 MB/s raw metadata
  • After fog aggregation: hourly summaries (1 update per hour per store) = 50 x 200 bytes/hour / 3600 = 2.8 bytes/second
  • Reduction: 1.5 MB/s to 2.8 bytes/s = 535,714x reduction at fog

Tier 3 - Cloud (central analytics):

  • 10 regional fog nodes x 2.8 bytes/s = 28 bytes/second
  • Daily data: 28 x 86,400 = 2.4 MB/day total for 500-store chain
  • Original raw: 324 TB/day
  • Overall reduction: 324 TB to 2.4 MB = 135 million x reduction

Cost comparison:

  • Cloud video storage: 324 TB/day x 30 days = 9,720 TB/month = 9,720,000 GB x $0.02/GB = $194,400/month
  • Metadata-only storage: 2.4 MB/day x 30 days = 72 MB/month = negligible cost
  • Edge inference hardware: 500 stores x $199 (Jetson Nano) = $99,500 one-time
  • ROI: Hardware pays for itself in 0.5 months (vs cloud video transmission/storage)

This example demonstrates that edge AI inference (person detection) combined with fog-tier aggregation makes large-scale video analytics economically viable – streaming 324 TB/day of raw video to the cloud would be cost-prohibitive.

55.10.1 Edge Data Reduction Calculator

Explore how different deployment scales affect data reduction ratios and cost savings across the three-tier architecture.

Requirement Edge (Device) Fog (Local Aggregation) Cloud (Centralized)
Latency needed <10ms 10-100ms >100ms acceptable
Scale (devices) 1-100 100-10,000 10,000+ (unlimited)
Compute complexity Simple rules, lightweight ML Multi-sensor fusion, regional analytics Deep learning training, cross-site correlation
Network dependency Must work offline Tolerates brief outages Requires connectivity
Data privacy Data stays on-premise Regional boundaries respected Data leaves site (compliance concerns)
Cost model CapEx (upfront hardware) Hybrid (gateway + connectivity) OpEx (pay-per-use)
Example platforms ESP32, Arduino, STM32 Raspberry Pi 4, NVIDIA Jetson AWS IoT Analytics, Azure IoT Hub
Best use case Airbag deployment, collision avoidance Factory floor dashboards, store analytics Predictive maintenance across global fleet

Selection heuristic:

  1. Start at the cloud, move left only if necessary: Begin with cloud analytics (easiest to develop/deploy). Move to fog if latency or bandwidth costs become prohibitive. Move to edge only when real-time control (<10ms) is safety-critical.
  2. Latency dictates tier: <10ms = edge, <100ms = fog, >100ms = cloud
  3. Scale determines hierarchy: <100 devices = cloud-only, 100-10K = fog+cloud, >10K = edge+fog+cloud
  4. Regulatory compliance: GDPR/HIPAA may mandate edge/fog for sensitive data (PII, health records)
Common Mistake: Deploying Edge AI Without Considering Model Staleness

Teams deploy ML models to edge devices (e.g., person detection on cameras, anomaly detection on sensors) but forget that models degrade over time as real-world conditions change. A person detection model trained on summer clothing fails in winter when everyone wears bulky coats.

What goes wrong: A smart factory deploys vibration-based anomaly detection on 100 machines, with models trained on 3 months of historical data. The models run locally on edge gateways for real-time alerting. Over 12 months, the factory gradually shifts production from aluminum parts (low vibration) to steel parts (higher baseline vibration). The anomaly detection model, still calibrated for aluminum, generates 50+ false alarms per day because normal steel vibration exceeds the aluminum-trained threshold.

Why it fails: Edge-deployed models are “frozen” – they do not adapt to changing conditions unless explicitly updated. Meanwhile, cloud models can retrain nightly on fresh data. The longer an edge model runs without updates, the more it drifts from reality.

The correct approach:

  1. Implement model versioning and over-the-air updates:

    import time
    # Edge device checks for model updates daily
    if time.time() - last_model_check > 86400:  # 24 hours
        new_model = check_server_for_update()
        if new_model.version > current_model.version:
            download_and_replace_model(new_model)
  2. Monitor model performance metrics at the edge:

    • Track prediction confidence scores over time
    • Alert if average confidence drops below threshold (indicates model drift)
    • Example: if person detection confidence was 95% in month 1 but 75% in month 6, model needs retraining
  3. Use hybrid edge-cloud approach for model training:

    • Edge: run inference with current model (low latency)
    • Cloud: collect edge inference logs + ground truth labels, retrain model weekly
    • Edge: receive updated model via OTA push

Real consequence: A facial recognition system for office access control deployed edge models on cameras at building entrances. Models were trained in 2023 and never updated. By 2025, employees wearing face masks (post-pandemic) had 40% false rejection rate because the model was not trained on masked faces. Security reverted to manual badge-only access, defeating the entire purpose of the biometric system. The fix: implement monthly model retraining using recent camera footage, with automated OTA deployment to all edge cameras. Model accuracy recovered to 98% within 2 months. The lesson: edge AI systems MUST include a model lifecycle management process – training, versioning, deployment, monitoring, and retraining – or models will degrade and fail over time.

55.11 See Also

Architecture Foundations:

Within This Module:

55.12 What’s Next

Current Next
Edge Architecture Review Edge Review: Calculations

Related topics:

Chapter Focus
Edge Review: Deployments Real-world deployment patterns and technology stack
Edge Topic Review Main review index