Edge Computing: What is it?

What is Edge Computing?

Edge computing is a data processing paradigm that moves part of the computation, storage, and software intelligence from centralised data centres to devices located at the "edge" of the network, as close as possible to the data source. Rather than sending all raw data to the cloud for processing, edge devices (Raspberry Pi, industrial gateways, local servers) filter, aggregate, and analyse data locally, only sending relevant results or alerts to the cloud.

This paradigm emerged from the convergence of two trends: the explosion in data volume generated by IoT sensors (making full cloud transfer impractical and costly) and the growing need for real-time responsiveness (where milliseconds of network latency to the cloud are unacceptable for critical applications). Edge computing does not replace the cloud: it complements it by creating a hybrid architecture where each layer processes data appropriate to its level.

In the IoT ecosystem, edge computing manifests at several levels: the "device edge" (processing directly on the sensor or microcontroller), the "gateway edge" (processing on the Raspberry Pi gateway that aggregates multiple sensors), and the "local edge" (on-site local server that processes data from multiple gateways before cloud transmission). Each level offers a different trade-off between computing power and data proximity.

Why Edge Computing matters

Edge computing addresses fundamental challenges that the cloud alone cannot solve, particularly in the IoT context where data is continuously generated by thousands of sensors.

• Ultra-low latency: local processing eliminates the network round-trip to the cloud (typically 50-200 ms), enabling reactions in milliseconds. Essential for security alerts, industrial control, and real-time anomaly detection.
• Bandwidth reduction: a vibration sensor sampling at 1 kHz generates megabytes of data per day. Edge processing extracts relevant metrics (dominant frequency, maximum amplitude) and transmits just a few bytes, reducing bandwidth by 99%.
• Network resilience: edge devices continue to operate and make decisions even during network outages, storing data locally for later synchronisation. Critical for remote sites with intermittent connectivity.
• Data sovereignty: GDPR imposes restrictions on personal data transfer. Edge computing allows data to be anonymised or aggregated locally before being sent to the cloud, facilitating regulatory compliance.
• Reduced cloud costs: less data transferred and stored in the cloud means reduced infrastructure bills, a significant advantage for large-scale IoT deployments.

How it works

The edge computing architecture in an IoT context is structured in layers. At the lowest level, sensors collect raw data (temperature every second, high-frequency vibrations, camera images). This data is transmitted to the edge node (typically a Raspberry Pi or embedded server) that executes the local processing logic.

Edge processing includes several types of operations. Filtering eliminates redundant or out-of-range data (no need to send 3,600 identical measurements per hour if the temperature has not changed). Aggregation computes statistics over time windows (average, min, max, standard deviation over 5 minutes). Anomaly detection compares real-time measurements against thresholds or learned models to trigger immediate alerts. ML inference runs lightweight machine learning models (TensorFlow Lite) for classification or prediction directly on the device.

Edge processing results are published via MQTT to the central backend (Django), which handles long-term storage, dashboards, advanced analytics, and cross-site coordination. The Raspberry Pi uses Python to implement edge logic, with libraries like NumPy for numerical processing, pandas for temporal aggregation, and paho-mqtt for communication. Docker containerises edge services to facilitate deployment and remote updates.

Concrete example

KERN-IT has approached edge computing from two sides. The first is research: Kenobi, a 24-month project carried out with the operator Venn Telecom and co-funded by Innoviris and VLAIO, aimed at a vendor-independent network management platform for ever larger and more heterogeneous SD-WAN, IoT and edge fleets. The edge part of the project consisted in designing an edge computer connected to the SD-WAN equipment and experimenting with its remote control: deploying and updating nodes without site visits, containerising and orchestrating processing on the edge, and deciding where intelligence should sit between the cloud and the field depending on the latency, bandwidth and reliability of the link. All of it was put to the test on more than 2,000 network points across Europe.

The second side is day-to-day operations: the self-service print kiosks KERN-IT has run for more than ten years for a university copy service are each a small edge computer. The kiosk identifies the user, reads the USB stick in an isolated space, starts the print job and counts pages locally; if the shop's network drops, it keeps serving and synchronises the till when the connection returns. The fleet, spread across several sites, is updated and monitored remotely, and a failed device is replaced by a new one configured from a recipe. These two experiences, research and field work, are what allow KERN-IT to say where intelligence belongs in a fleet of devices, and where it does not.

Implementation

1. Identify processing to offload: analyse which operations benefit from local execution (real-time alerts, filtering, aggregation) versus those requiring the cloud (historical analysis, dashboards, advanced ML).
2. Choose edge hardware: Raspberry Pi 4/5 for standard projects, NVIDIA Jetson for ML inference on images/video, local server for sites with many gateways.
3. Develop the processing pipeline: implement filtering, aggregation, and anomaly detection in Python on the Raspberry Pi, with configurable time windows.
4. Implement offline buffering: add local storage (SQLite, JSON files) for data awaiting synchronisation, with an automatic retry mechanism after reconnection.
5. Containerise services: package the edge pipeline in Docker containers for reproducible deployment and remote updates via Fabric or Ansible.
6. Monitor edge nodes: track the health of Raspberry Pi devices (CPU, memory, disk space, temperature) via metrics published on MQTT and displayed on the central dashboard.

Associated technologies and tools

• Raspberry Pi: the most widely used edge computing platform in IoT, offering a full Linux OS and GPIO interfaces for sensors.
• Docker: edge service containerisation for uniform deployment and simplified updates across device fleets.
• MQTT: lightweight messaging protocol for communication between edge nodes and the central backend.
• TensorFlow Lite: lightweight version of TensorFlow for running ML model inference directly on edge devices.
• Python (NumPy, pandas): language and libraries for numerical processing and data aggregation on edge nodes.
• SQLite: embedded database used as a local buffer for offline storage of data awaiting synchronisation.

Conclusion

Edge computing transforms IoT architectures by bringing intelligence closer to the field, reducing latency, bandwidth, and cloud dependency while improving resilience and GDPR compliance. It is not an alternative to the cloud but an essential complement that optimises the entire processing chain. At KERN-IT, we implement edge computing on Raspberry Pi at the heart of our IoT solutions, combining local Python processing with our Python backend (Flask or Django depending on the project) and KERN MAP visualisation to deliver responsive, reliable, and cost-effective monitoring systems to our Belgian clients.