Examine the constraints
When connectivity is unreliable and a quick response is needed, some data is processed close to its source. A central service can combine information from many sites. The choice depends on latency, connectivity and the work that must continue if the link fails.
One object, several tasks
Imagine a production-line camera. It must detect an obstruction quickly, preserve an event for investigation and send a summary to analytics spanning several sites. These tasks have different sensitivities to delay and lost connectivity. They need not all compute in one place. Urgent local response, storage and shared analysis can be separated according to outcome requirements.
First record what must happen without a network: whether safe work can continue, how many events to store and which functions to suspend. Then determine whether the central service needs the full stream or selected events. Sending only what is needed reduces bandwidth but limits later reanalysis. Make that trade-off consciously for the task.
Local computation needs infrastructure too
On-site devices must be installed, updated, monitored and replaced. Power, cooling and compute are limited. Behaviour-changing model updates need testing before broad installation. Cloud removes some hardware responsibilities but adds channel, external-service and billing dependencies. A demonstration request's speed is insufficient for comparison.
Distributed systems need version compatibility. Some devices may update while others remain offline. Central analytics must know which rules produced events and whether they are comparable, especially when a new model identifies the same object differently. Processing version and event time become useful data rather than mere developer details.
Tests should include failure
Simulate a lost channel, device restart and full local storage during the pilot. Observe operator information, lost events and recovery. Check ordering too: a late-delivered event must not silently appear new. Such trials teach more than comparing ideal scenarios with stable connections and empty queues.
