Here's something worth checking during your next architecture review: has anyone on the team documented what "real-time" actually means in milliseconds for your specific use case?
Because cloud vendors and control system engineers use completely different definitions, this gap causes real problems in production.
Cloud platforms reliably deliver sub-second responses. For dashboards and reporting, that's fine. But for manufacturing automation, safety monitoring, or fraud detection at transaction speed, sub-second is often 50x too slow.
A production line running 400 units per minute has a 150ms decision window per unit. A typical cloud round-trip takes 300 to 500ms. By the time the system responds, 3 to 4 units have already passed the inspection point.
That's not a performance issue you can optimise away. Fibre propagation, routing hops, serialisation, TLS, ingestion queuing; these add up before a single line of your analytics code runs.
The practical breakdown:
- **Sub-second: **Cloud works well. Dashboards, reporting, alerting.
- Sub-100ms: Cloud can work, but needs careful design and testing under real network conditions.
- Sub-10 to 20ms: Edge or on-premises only. Cloud cannot reliably serve this tier.
Most teams working on control systems, robotics, or safety-critical monitoring need the third tier. Worth getting that defined before the architecture is locked in.
If you want the full latency breakdown, including the physics behind why optimisation can't close the gap, it's all here: https://dev.to/actiandev/why-real-time-analytics-cant-depend-on-cloud-in-2026-1paj
