Virtualization lets you ignore which physical server a VM runs on — until you can't. The hardware still sits beneath the abstraction, and its behavior leaks through. Two identically-specified cloud instances can perform very differently. The cause is shared tenancy: a given instance may sit on degraded hardware, or share a host with someone else's resource-hungry workload (the "noisy neighbor"). The API presents a uniform facade; the silicon underneath does not honor it. The practical consequence is that you cannot treat the cloud as infinite, uniform capacity. Reasoning about the physical layer pays off: - **Instance sizing** — some teams pick a VM size large enough to get a host entirely to themselves, even when they don't need the full capacity, to avoid noisy neighbors. - **Storage and networking** — the fastest option (e.g. local SSD) is not automatically the right one; it carries tradeoffs in portability, cost, and availability. This is mechanical sympathy applied to the cloud: the deeper your understanding down the stack into the hardware, the more reliably and efficiently your systems run. ## Source - [Mechanical Sympathy and the Cloud](https://infrastructure-as-code.com/posts/mechanical-sympathy.html) — Kief Morris, 23 Mar 2015