The assumption that all computation should happen in centralised cloud data centres is being challenged by applications that need real-time responses, operate in low-connectivity environments, or generate data volumes that are impractical to transmit to the cloud. Edge computing — processing data near its source — addresses these constraints.
When Edge Computing Makes Sense
Edge computing adds architectural complexity, so it should be adopted deliberately rather than as a default. The strongest use cases share one or more characteristics: latency requirements below what cloud round-trips can achieve (autonomous vehicles, industrial control systems), bandwidth constraints that make transmitting raw data impractical (video analytics, sensor arrays), privacy requirements that prohibit sending data to external systems, or reliability requirements that demand operation during connectivity outages.
For applications without these constraints, cloud-based processing is simpler, cheaper, and easier to maintain. The decision to deploy edge computing should be driven by specific technical requirements, not architectural fashion.
Architecture Patterns
Edge architectures typically follow a tiered model. The device tier handles immediate data collection and simple filtering. The edge tier — local servers or gateways — handles real-time processing, aggregation, and local decision-making. The cloud tier handles long-term storage, model training, fleet management, and analytics that benefit from centralised data.
The critical design decision is what to process at each tier. Edge devices have limited compute and storage, so processing must be selective. The general principle is: process at the edge what must be processed in real time, and send everything else to the cloud for deeper analysis.
Edge AI and ML Inference
Running machine learning models at the edge enables real-time inference without cloud connectivity. Optimised model formats — TensorFlow Lite, ONNX Runtime, and hardware-specific compilers — allow complex models to run on edge hardware with acceptable latency. The trade-off is typically between model accuracy and inference speed: smaller, faster models may sacrifice some accuracy compared to their cloud-hosted counterparts.
Model lifecycle management at the edge introduces challenges that cloud deployments avoid. Updating models across thousands of edge devices, monitoring model performance in diverse environments, and handling graceful degradation when models encounter out-of-distribution inputs all require infrastructure that does not exist in standard cloud deployment pipelines.
Security at the Edge
Edge devices operate outside the physical security of data centres, making them vulnerable to physical tampering, network attacks, and firmware exploitation. Hardware security modules, secure boot processes, encrypted storage, and regular security updates are baseline requirements. The attack surface of an edge deployment is proportional to the number of deployed devices, making security automation essential at scale.
