Cloud or edge? Place the workload, not the slogan

Follow one transaction from start to finish

Describe the full transaction: sensing, uplink, queue, compute, storage and response. In an illustrative 50 ms deadline, an 18 ms uplink, 12 ms queue, 15 ms computation and 8 ms return total 53 ms: faster processors alone may not repair the design. These are assumed values, not field measurements. Measure tail latency under load and distinguish firm control deadlines from interactive user experience. Moving one component locally helps only if its dependencies also fit the budget.

Give the site a plan for losing the cloud

Separate the data plane from the control plane. A site may continue approved local inference during loss of central orchestration, while deferring model updates and nonessential analytics. Define what expires, which credentials remain valid, how logs are buffered and when the system must fail safely. Data residency requires mapping copies, backups and remote support paths; placing a server in one city does not by itself constrain every copy.

Test the operating model, including recovery

Evaluate three options: centralized, local and hybrid. Compare useful availability, bandwidth, local maintenance, update rollback and total operating effort. Replication is not a backup when corruption propagates to every replica.

Reconcile offline writes with explicit conflict rules and test a prolonged link outage. For a Sakarya industrial site, the engineering brief should describe actual access circuits, local power and operational response. A regional label is not a substitute for measurements or a promise of on-site capacity.