45 Edge Processing Patterns
analytics-ml
edge
patterns
processing
45.1 Start With the Story
Picture an IoT team using the ideas in Edge Processing Patterns during a live operations review. A device has produced messy evidence, an analytic step is about to change an alert or control decision, and someone has to explain why the result should be trusted.
Read this page as that path from sensor evidence to accountable action. Start with what the system observes, keep the model or data treatment visible, and finish with the check that would convince an operator, maintainer, or auditor to act.
Chapter Roadmap
This chapter turns patterns into placement decisions:
- First separate Filter, Aggregate, Infer, and Store-Forward by the operational problem each one solves.
- Then test threshold filtering, aggregation, inference, and outage buffering.
- Next compare edge ML, cloud ML, batch, streaming, and cost breakpoints.
- Finally apply the pattern mix to the Danfoss refrigeration case and practice quizzes.
Checkpoints recap the logic; optional calculators can wait.
45.2 Learning Objectives
By the end of this chapter, you will be able to:
- Apply Four Edge Processing Patterns: Implement Filter, Aggregate, Infer, and Store-Forward patterns for different IoT scenarios
- Select Optimal Patterns: Match processing patterns to application requirements based on latency, bandwidth, and reliability needs
- Evaluate Trade-offs: Compare edge ML inference versus cloud ML, and batch processing versus real-time streaming
- Avoid Common Mistakes: Understand when edge processing saves costs versus when it adds unnecessary complexity
Key Concepts
- In-network processing: Executing data transformations within the network infrastructure (on switches, gateways, or routers) rather than at endpoints, reducing end-to-end latency and backbone bandwidth usage.
- Stream operator: A discrete processing step in a streaming pipeline (filter, map, aggregate, join) that transforms one stream into another; composing stream operators builds a complete edge processing pipeline.
- Event windowing: Grouping stream events into finite sets for batch processing based on time (tumbling windows, sliding windows) or event count (count windows) before applying aggregation or detection operators.
- Stateful processing: Stream processing that maintains state between events (running averages, counters, session trackers) in contrast to stateless operations (format conversion, field extraction) that treat each event independently.
- Processing latency budget: The maximum time allowed for all edge processing steps combined before data must be acted upon or forwarded, constraining the complexity and number of processing operators that can be applied.
In-network processing can be organised as a sense-and-execution pipeline rather than a single gateway script. Sensors, switches, routers, gateways, and actuators each hold part of the decision path: classify the incoming record, combine it with nearby context, choose whether the result belongs locally or upstream, and execute only the bounded action. Body-and-brain architectures use that split deliberately: the body performs local sensing and actuation, while the brain coordinates policy, learning, and cross-site optimisation.
For Beginners: Edge Data Processing
Edge data processing means analyzing IoT data right where it is collected, before sending anything to the cloud. Think of a store manager who handles routine customer questions on the spot instead of calling headquarters for every decision. This makes your IoT system faster, cheaper, and more reliable.
45.3 Prerequisites
Before diving into this chapter, you should be familiar with:
- Edge Compute Patterns Overview: Introduction to edge computing concepts
- IoT Reference Model: Understanding the seven-level architecture and Level 3 processing
Use those prerequisites as the frame: what stays local, and what evidence moves upstream?
45.4 Edge Processing Patterns
Edge computing employs four primary data processing patterns, each optimized for different IoT scenarios:
Filter
Send only threshold alerts.
Aggregate
Send summaries instead of raw streams.
Infer
Run ML locally and send anomalies.
Store-Forward
Buffer data during outages and sync later.
Mobile-friendly summary of the four core edge processing patterns.
Latency-Based Processing Location
This view helps decide whether a processing step belongs on the edge device, an intermediate fog gateway, or the cloud before selecting the specific pattern:
Match latency, compute, and bandwidth constraints to the processing location, then choose Filter, Aggregate, Infer, or Store-Forward inside that tier.
Edge Processing Timeline
This view shows how data flows through edge processing over time:
Edge processing pipelines execute in milliseconds while cloud sync happens asynchronously.
45.5 Pattern Selection Guide
| Pattern | Best For | Bandwidth Savings | Use Case Example |
|---|---|---|---|
| Filter | Threshold monitoring | 99%+ | Send alerts only when temperature exceeds safe limits |
| Aggregate | Trend analysis | 99%+ | Send hourly averages instead of per-second readings |
| Infer | Anomaly detection | 95%+ | Visual inspection - send defect alerts, not all images |
| Store & Forward | Intermittent connectivity | N/A | Remote sites with satellite links - sync when online |
Checkpoint: Choose the Pattern
- Filter and Aggregate both aim for 99%+ bandwidth savings, but Filter preserves exception timing while Aggregate preserves trends.
- Infer fits local ML decisions when edge latency matters more than cloud accuracy.
- Store-Forward is the reliability pattern when the link is intermittent rather than merely expensive.
Read each pattern as a contract: what the gateway keeps, forwards, and can reconstruct.
45.5.1 Pattern 1: Filter at Edge
The filter pattern applies simple threshold checks to determine which data warrants transmission:
# Example: Temperature threshold filter
def filter_reading(temp_celsius, threshold=80.0):
if temp_celsius > threshold:
return {"alert": True, "value": temp_celsius, "timestamp": now()}
return None # Don't transmit normal readingsBest for:
- Alarm/alert systems where only exceptions matter
- High-frequency sensors where most readings are routine
- Bandwidth-constrained links where every byte counts
45.5.2 Pattern 2: Aggregate at Edge
The aggregate pattern computes statistics locally and sends summaries:
# Example: Statistical aggregation
import statistics
def aggregate_window(readings, window_minutes=60):
return {
"min": min(readings),
"max": max(readings),
"avg": sum(readings) / len(readings),
"stddev": statistics.stdev(readings),
"count": len(readings),
"window_end": now()
}
Putting Numbers to It
Consider 200 temperature sensors sampling at 1 Hz:
Without aggregation (transmit every reading): \[\text{Per sensor} = 1\text{ Hz} \times 4\text{ bytes} = 4\text{ bytes/s}\] \[\text{Fleet total} = 4\text{ bytes/s} \times 200 = 800\text{ bytes/s}\] \[\text{Daily bandwidth} = 800\text{ bytes/s} \times 86{,}400\text{ s} = 69.1\text{ MB/day}\]
With 1-hour aggregation (transmit min/max/avg/stddev every hour): \[\text{Per sensor/hour} = 4\text{ values} \times 4\text{ bytes} = 16\text{ bytes}\] \[\text{Per sensor/day} = 16\text{ bytes} \times 24\text{ hours} = 384\text{ bytes}\] \[\text{Fleet daily bandwidth} = 384\text{ bytes} \times 200 = 76{,}800\text{ bytes} = 75\text{ KB/day}\]
Reduction ratio: \[\frac{69.1\text{ MB} - 0.075\text{ MB}}{69.1\text{ MB}} = 99.89\% \text{ reduction}\]
Aggregation compresses 3,600 readings into 4 statistical values, achieving near-perfect bandwidth savings while preserving trend information for analytics.
Best for:
- Trend analysis where individual readings are less important than patterns
- Environmental monitoring (temperature, humidity, air quality)
- Capacity planning and historical analysis
45.5.3 Pattern 3: Infer at Edge
The infer pattern runs machine learning models locally:
# Example: Anomaly detection at edge
model = load_tflite_model("anomaly_detector.tflite")
def infer_anomaly(sensor_data):
prediction = model.predict(sensor_data)
if prediction["anomaly_score"] > 0.85:
return {"anomaly": True, "score": prediction["anomaly_score"],
"features": sensor_data}
return None # Normal - don't transmitBest for:
- Visual inspection (defect detection in manufacturing)
- Predictive maintenance (vibration analysis)
- Safety systems requiring immediate response
45.5.4 Pattern 4: Store and Forward
The store-and-forward pattern handles intermittent connectivity:
# Example: Store-and-forward buffer
class EdgeBuffer:
def __init__(self, max_size_mb=100):
self.buffer = []
self.max_size = max_size_mb * 1024 * 1024
def store(self, reading):
self.buffer.append(reading)
self.compact_if_needed()
def forward_when_connected(self):
if network_available():
batch = self.buffer.copy()
if upload_batch(batch):
self.buffer.clear()Best for:
- Remote sites (oil rigs, agricultural fields, offshore platforms)
- Mobile assets (vehicles, shipping containers, drones)
- Any deployment with unreliable connectivity
Checkpoint: Run the Four Contracts
- A threshold filter sends exceptions and discards routine readings.
- A windowed aggregate compresses many samples into min, max, avg, stddev, count, and window metadata.
- An inference step sends anomaly evidence, while Store-Forward keeps buffered readings until the network returns.
Next, decide which trade-off the deployment can defend when latency, accuracy, power, connectivity, and cost conflict.
45.6 Trade-off Analysis
45.6.1 Edge ML Inference vs Cloud ML Inference
Edge ML vs Cloud ML Inference
Option A (Edge ML): Deploy 2-10 MB quantized models on gateways (TensorFlow Lite, ONNX Runtime), achieving 10-50 ms inference latency with 85-92% accuracy for classification tasks on Cortex-M4/ESP32 class devices.
Option B (Cloud ML): Run 100 MB - 1 GB full-precision models in cloud (TensorFlow Serving, AWS SageMaker), achieving 200-500 ms round-trip latency with 95-99% accuracy using GPUs for complex pattern recognition.
Decision Factors: Choose edge ML when latency requirements are under 100 ms (safety systems, real-time control), privacy mandates data cannot leave premises (healthcare, industrial IP), or connectivity is unreliable (remote assets, mobile equipment). Choose cloud ML when model complexity requires GPU acceleration (video analytics, NLP), training data continuously improves models (recommendation systems), or centralized management simplifies updates. Hybrid architectures run simple detection at edge (anomaly flags, threshold checks) while cloud handles deep analysis (root cause diagnosis, long-term forecasting). A factory safety system needs 20 ms edge response, but weekly predictive maintenance reports can use cloud-trained models updated monthly.
45.6.2 Batch Processing vs Real-Time Streaming
Batch vs Streaming at Edge
Option A (Batch Processing): Collect sensor data locally for 1-60 minutes, process in batches, upload aggregated results. Reduces compute cycles by 80-95%, extends battery life 3-5x on constrained devices, but introduces 1-60 minute detection latency.
Option B (Real-Time Streaming): Process each sensor reading immediately as it arrives with sub-second latency. Enables instant anomaly detection and immediate control responses, but requires 5-10x more edge compute power and continuous network connectivity for cloud integration.
Decision Factors: Choose batch processing for delay-tolerant analytics (hourly environmental reports, daily asset utilization), bandwidth-constrained links (satellite, cellular metered), and battery-powered devices where duty cycling extends deployment life from weeks to years. Choose real-time streaming for safety-critical monitoring (gas leaks, machine failures requiring less than 1 s response), control systems with tight feedback loops (HVAC, robotics), and applications where stale data has no value (live tracking, interactive systems). A smart meter can batch hourly readings, but a pipeline pressure sensor must stream in real-time to detect ruptures within seconds.
45.7 When Edge Saves Money
Edge Does Not Always Save Cost
The Myth: “Processing at the edge always saves money compared to cloud processing.”
The Reality: Edge computing reduces bandwidth costs but introduces hardware and maintenance costs that can exceed cloud savings in many scenarios.
Real-World Example: Agricultural Soil Monitoring
A precision agriculture company deployed 10,000 soil moisture sensors across 5,000 acres:
Cloud-Only Approach (Initial Design):
- 10,000 sensors x 1 reading/hour x 24 hours x 30 days = 7.2M readings/month
- Data size: 7.2M readings x 50 bytes = 360 MB/month
- Cloud egress cost: 360 MB / 1,024 MB/GB x $0.09/GB = $0.03/month
- Cloud compute/storage: ~$50/month
- Total: ~$50/month
Edge Gateway Approach (Actual Deployment):
- 50 edge gateways ($200 each): $10,000 upfront capital
- Edge aggregation reduces cloud traffic by 90%: 36 MB/month
- Cloud costs: $5/month (minimal compute)
- Gateway maintenance: $100/month (cellular data, power, repairs)
- Amortized gateway cost over 3 years: $278/month
- Total: $383/month
The Hidden Costs:
- Hardware depreciation: $10,000 / 36 months = $278/month
- Cellular connectivity: 50 gateways x $2/month = $100/month
- Maintenance visits: 1 failed gateway/month x $150/visit = $150/month (later reduced with better hardware)
- Software updates: Edge devices require OTA update infrastructure
When Edge Actually Saved Money (Year 2): After initial deployment issues were resolved and maintenance costs dropped to $50/month:
- Edge total: $278 (amortized) + $100 (cellular) + $50 (maintenance) + $5 (cloud) = $433/month
- Cloud total: $50/month
Edge remained more expensive until the company expanded to 100,000 sensors in Year 3. At that scale, cloud compute costs grew to ~$500/month while edge gateways were already provisioned with spare capacity, making the per-sensor cost of edge processing lower than cloud.
Key Takeaway: Edge computing provides the greatest cost savings when:
- High data volumes overwhelm bandwidth costs (>10 GB/month)
- Hardware is amortized over multi-year deployments
- Maintenance is minimal (reliable hardware, remote updates)
- Latency/privacy requirements justify the investment regardless of cost
Decision Framework:
- Small deployments (<1,000 sensors, <1 GB/month): Cloud is usually cheaper
- Medium deployments (1,000-10,000 sensors): Hybrid (edge aggregation + cloud) often optimal
- Large deployments (>10,000 sensors, >10 GB/month): Edge pays for itself within 6-12 months
Checkpoint: Price the Trade-off
- Edge ML is attractive for 10-50 ms decisions, while cloud ML often pays a 200-500 ms round trip.
- Batch processing can reduce compute cycles by 80-95%, but it introduces 1-60 minute detection latency.
- Cost savings need enough volume: the chapter’s decision framework separates small deployments, medium deployments, and large deployments above 10,000 sensors or 10 GB/month.
Those trade-offs are clearest in a fleet where each stream can choose its own pattern.
45.7.1 Danfoss Refrigeration Edge
Scenario: Danfoss, a Danish climate technology company, manages refrigeration controllers in 40,000 European supermarkets. Each store has 12 display cases with 3 sensors each (temperature, defrost status, compressor current), reporting every 30 seconds.
Given:
- 40,000 stores x 12 cases x 3 sensors = 1,440,000 sensors
- 30-second intervals = 86,400 s / 30 s = 2,880 readings/sensor/day
- Raw payload: 8 bytes per reading (timestamp + value)
- Cellular backhaul at EUR 0.50/MB
Step 1: Calculate raw data volume
| Metric | Value |
|---|---|
| Daily readings (total) | 1,440,000 x 2,880 = 4.15 billion |
| Daily raw data | 4.15 x 10^9 x 8 bytes = 33.2 GB |
| Daily cellular cost | 33,200 MB x EUR 0.50/MB = EUR 16,600 |
Step 2: Apply edge patterns per sensor type
| Sensor | Pattern | Logic | Reduction |
|---|---|---|---|
| Temperature | Filter | Transmit only if outside -22 to -18 C (freezer) or 2 to 4 C (chiller) | 97% (normal 97% of time) |
| Defrost status | Filter | Transmit only on state change (start/stop) | 99.5% (2 events/day vs 2,880) |
| Compressor current | Aggregate | Transmit 5-minute RMS + peak | 90% (10 summaries/hour vs 120 raw) |
Step 3: Calculate optimized volumes
Each sensor type accounts for one-third of the 1,440,000 sensors (480,000 sensors each):
| Metric | Raw | After Edge | Savings |
|---|---|---|---|
| Temperature data/day | 11.1 GB | 333 MB | 97% |
| Defrost data/day | 11.1 GB | 55 MB | 99.5% |
| Compressor data/day | 11.1 GB | 1.1 GB | 90% |
| Total daily | 33.2 GB | 1.49 GB | 95.5% |
| Cellular cost/day | EUR 16,600 | EUR 745 | EUR 15,855 saved |
Result: Edge processing reduces daily cellular costs from EUR 16,600 to EUR 745 – a 95.5% reduction. The fleet-wide gateway investment (40,000 gateways x EUR 120 = EUR 4.8 million) recovers in about 10 months through cellular savings of EUR 15,855 per day.
Key Insight: The three sensor types in the same deployment use three different edge patterns. Temperature uses Filter (only exceptions matter), defrost uses Filter (only state changes matter), and compressor current uses Aggregate (trend data matters). Pattern selection is per-sensor, not per-deployment.
Checkpoint: Validate the Fleet Mix
- The Danfoss case starts with 40,000 stores, 12 display cases per store, and 3 sensors per case.
- The raw daily total is 33.2 GB, and the mixed edge plan reduces it to 1.49 GB.
- The same calculation drops daily cellular cost from EUR 16,600 to EUR 745, but only because each sensor type gets the pattern that matches its evidence need.
45.8 Continue: Edge Stream Window Contracts
The four processing patterns above still need precise stream semantics before a gateway emits reliable metrics. Continue to Edge Stream Window Contracts for tumbling, sliding, and session window rules, state budgets, late-event policy, and replay metadata.
45.9 Summary
- Four edge processing patterns address different IoT requirements: Filter (threshold alerts), Aggregate (statistical summaries), Infer (ML-based detection), and Store-Forward (intermittent connectivity)
- Pattern selection depends on primary priority: bandwidth reduction, real-time response, reliability, or privacy
- Edge ML trade-offs balance latency (10-50 ms edge vs 200-500 ms cloud) against accuracy (85-92% edge vs 95-99% cloud)
- Batch vs streaming trade-offs balance power efficiency against detection latency
- Cost analysis shows edge saves money only at scale (>10,000 sensors) or when latency/privacy requirements justify hardware investment
- Hybrid architectures typically provide the best balance: edge for time-critical decisions, cloud for complex analytics
Key Takeaway
The four edge processing patterns – Filter, Aggregate, Infer, and Store-Forward – each solve a different problem. Filter reduces bandwidth by sending only exceptions. Aggregate computes local statistics for trend analysis. Infer runs ML models for intelligent detection. Store-Forward handles intermittent connectivity. Hybrid architectures combining edge speed with cloud analytics consistently outperform either approach alone, but edge only saves money at scale (>10,000 sensors) or when latency and privacy requirements justify hardware investment.
45.10 Concept Check
45.11 Concept Relationships
How This Concept Connects
Builds on:
- Edge Compute Patterns Overview - Foundation concepts
- Edge IoT Reference Model - Level 3 processing context
Four Patterns in Detail:
- Filter Pattern - 99%+ bandwidth reduction by sending only threshold violations
- Aggregate Pattern - Statistical summaries (min/max/avg) for trend analysis
- Infer Pattern - ML models run locally, send anomaly alerts only
- Store-Forward Pattern - Buffers data during outages, syncs when reconnected
Real-World Examples:
- Danfoss Supermarkets - 95.5% data reduction across 40,000 stores using three different patterns
- Agricultural Soil Monitoring - Cost analysis shows cloud-only cheaper until very large scale
45.12 See Also
Related Resources
Pattern Implementation:
- Edge Patterns Practical Guide - Interactive tools and worked examples
- Edge Cyber-Foraging - Opportunistic compute offloading
Architecture Context:
- Edge, Fog, and Cloud Overview - Three-tier architecture
- Multi-Sensor Data Fusion - Combining patterns for sensor fusion
Case Studies:
- Danfoss Case Study (in chapter) - Three patterns for three sensor types
- Agricultural Cost Analysis (in chapter) - When edge investment does not pay off
45.13 Try It Yourself
Four Edge Patterns Exercise
Scenario: Industrial motor monitoring with vibration sensors at 1 kHz sampling.
Pattern 1: Filter
def filter_pattern(vibration_g, threshold=5.0):
"""Send only threshold violations"""
if vibration_g > threshold:
return {"alert": True, "value": vibration_g, "pattern": "filter"}
return None # Don't transmit normal readings
# Test: 8 samples, 3 exceed threshold (5.0)
readings = [2.1, 3.5, 2.8, 6.2, 4.1, 7.8, 3.2, 5.9]
alerts = [filter_pattern(r) for r in readings if filter_pattern(r)]
print(f"Filter: {len(readings)} readings -> {len(alerts)} alerts")Pattern 2: Aggregate
import statistics
def aggregate_pattern(readings_1sec):
"""Compute 1-second statistics from 1000 Hz samples"""
return {
"min": min(readings_1sec),
"max": max(readings_1sec),
"avg": statistics.mean(readings_1sec),
"stddev": statistics.stdev(readings_1sec),
"pattern": "aggregate"
}
# Test: 1000 samples -> 4 values = 250x reduction
samples_1khz = [2.1 + (i * 0.01) for i in range(1000)]
summary = aggregate_pattern(samples_1khz)
print(f"Aggregate: 1000 samples -> 4 values = 250x reduction")Pattern 3: Infer (simplified)
def infer_pattern(vibration_rms, threshold_trained=4.5):
"""Simplified ML inference: anomaly score based on trained threshold"""
anomaly_score = vibration_rms / threshold_trained
if anomaly_score > 1.2: # 20% above normal
return {"anomaly": True, "score": round(anomaly_score, 2),
"pattern": "infer"}
return None # Normal operation
# Test: 5.8 g RMS is above trained baseline of 4.5 g
result = infer_pattern(5.8)
print(f"Infer: {result}")Pattern 4: Store-Forward
class StoreForwardBuffer:
def __init__(self, max_size_mb=100):
self.buffer = []
self.max_size = max_size_mb * 1024 * 1024
def store(self, reading):
self.buffer.append(reading)
if len(self.buffer) > 1000: # Simplified FIFO
self.buffer.pop(0)
def forward(self, network_available):
if network_available and self.buffer:
print(f"Forwarding {len(self.buffer)} buffered readings")
self.buffer.clear()
return True
return False
# Test: buffer during outage, then forward
buffer = StoreForwardBuffer()
for i in range(50):
buffer.store({"reading": i})
buffer.forward(network_available=False) # No sync
print(f"Buffer size: {len(buffer.buffer)} readings")
buffer.forward(network_available=True) # Sync successful
print(f"After sync: {len(buffer.buffer)} readings")What to Observe:
- Filter drastically reduces transmissions (99%+) but loses trend information
- Aggregate preserves statistics while achieving 250x reduction
- Infer requires trained model but enables intelligent detection
- Store-Forward ensures no data loss during outages
Extension Challenge: Combine patterns: Filter + Aggregate. Send hourly aggregates for normal operation, immediate alerts for threshold violations.
45.14 What’s Next
| Next Topic | Description |
|---|---|
| Edge Stream Window Contracts | Window semantics, state budgets, and replay metadata for edge processing outputs |
| Cyber-Foraging and Caching | Opportunistic compute offloading and caching strategies |
| Edge Patterns Practical Guide | Interactive tools, worked examples, and common pitfalls |
