https://medium.com/digital-catapult/distributing-a-machine-learning-algorithm-across-iot-device-edge-and-cloud-731480bfcceb

When designing a new IoT system, there are many tradeoffs which determine the best way to distribute a machine learning algorithm across the different components of the system: device, edge and cloud. Battery lifetime, physical size, cost, real-time connectivity needs, privacy concerns and debug/troubleshoot needs are just some of the issues a system architect needs to consider when designing a system.

Typical IoT system architectures include devices (or nodes) which are deployed in physical spaces and usually include one or more sensors; hubs (or gateways or edge) which bridge between communication protocols and are located relatively close to the devices; a centralised cloud environment which stores and processes the data, and front-end devices which users can interact with, explore the data and get notifications. Obviously, there are scenarios where the devices communicate directly with the cloud environment and scenarios where the devices act as the frontend device, but logically, this architecture describes the common roles in the setup.