Google Compute Engine

You need to run an application long-term, but buying physical servers means high upfront costs and weeks-long lead times for scaling. If your traffic predictions are wrong, you either waste money or your service slows down.

Before cloud VMs, getting a new server meant a capacity-planning cycle that could stretch months: forecast traffic, write a purchase order, wait for hardware to arrive, rack and cable it, install an OS, configure networking. All of it was capital expenditure tied to a three-to-five-year hardware refresh cycle. If a traffic spike hit before the next refresh, you had no recourse — the fixed capacity was all you had, and users hit errors or slow responses until the spike passed. Over-provisioning to avoid that risk meant paying for idle hardware year-round. Cloud VMs replaced this with minute-level provisioning: request an instance, get it running, and delete it when the load subsides.

When a request to create an instance arrives, the hypervisor on a physical host carves out the specified CPU and memory resources and boots the VM with your chosen machine type and OS image. The instance lands inside a VPC with the firewall rules you defined. For production workloads, a Managed Instance Group watches load metrics — CPU utilization, request count, or a custom signal — and compares them against the threshold you set. When the metric crosses the threshold, new instances are added; when it drops, excess instances are removed. Persistent Disk survives instance deletion because it is network-attached storage, not a local drive — so you can detach it and reattach it to a different instance without losing data.

Both Compute Engine and Cloud Run run applications, but Compute Engine gives you direct VM management while Cloud Run handles infrastructure automatically once you deploy a container. Choose Compute Engine when you need OS-level control, custom drivers, or long-running processes.

Ideal for web servers, long-running batch jobs, and stateful applications that need to stay up continuously. Not the right fit when you only need to run code briefly in response to incoming requests.