Latency Is Physics

The speed of light does not negotiate. Distance between your workload and its data has a floor, and no cloud provider can undercut it.

The Constraint That Cannot Be Engineered Away

Light travels roughly 200 kilometres per millisecond through fibre — slower than vacuum, because glass. A round trip from London to a regional cloud datacentre 150 km away costs at minimum 1.5 ms in propagation alone, before queuing, serialisation, or any software overhead. Put your application in one AWS region and its primary database in another, and you have bought latency that no amount of tuning recovers.

This matters for a specific and identifiable class of workload. High-frequency trading systems, real-time control loops, synchronous distributed transactions, and multiplayer game servers with tight tick rates all have latency budgets measured in single-digit milliseconds. For them, a regional cloud endpoint is not a substitute for local infrastructure — it is a different category of thing. The managed database bargain includes this constraint in fine print most teams read too late: the convenience of a fully managed service means you accept wherever the provider places it.

Latency also compounds. A microservices chain making eight synchronous calls, each crossing even a modest distance, accumulates those floors. The application does not feel slow because the code is inefficient; it feels slow because geography is expensive.

The workloads genuinely unbothered by this are, in fact, most of them — batch processing, asynchronous pipelines, document storage, analytics. For those, a 20 ms round trip is invisible. The mistake is not moving latency-sensitive work to the cloud; the mistake is moving it there without first measuring what the propagation floor actually is, and checking whether the application can tolerate it.

Measure first. Physics does not offer a refund.