Maximizing Cost Efficiency with Preemptible VMs in GCP

Preemptible VM instances exist to provide a cost-effective solution for workloads that can tolerate interruptions. With discounts of up to 91% compared to standard VMs, they are ideal for batch processing, testing, and other non-critical tasks. However, their low cost comes with the caveat of potential preemption, which can disrupt your operations if not managed properly.

When a preemptible VM is at risk of being terminated, Compute Engine sends a preemption notice in the form of an ACPI G2 Soft Off signal. This notice gives you a brief window—up to 30 seconds—to execute a shutdown script and perform any necessary cleanup before the instance is forcibly stopped. If the instance does not stop within this period, it receives an ACPI G3 Mechanical Off signal, transitioning it to a TERMINATED state. It's crucial to design your applications to handle this preemption gracefully to avoid data loss or incomplete tasks.

In production, you need to be aware of several limitations. Preemptible instances cannot live migrate to regular VMs, nor can they automatically restart during maintenance events. They are also not covered by any Service Level Agreement (SLA), meaning you cannot rely on them for critical workloads. If a preemptible instance is stopped less than one minute after creation, you won't be billed for it, which can be a silver lining for testing. Additionally, you won't incur charges for Local SSD disk usage or GPUs if the instance is preempted shortly after starting. Understanding these nuances will help you leverage preemptible VMs effectively while avoiding pitfalls.