Preempt and Resume

Preempt and Resume | Generated by AI

Home 2026.07

Preempt = to interrupt or take priority over something that’s currently running, in order to do something else first.

Common uses:

OS/scheduling: A higher-priority process interrupts a lower-priority one running on the CPU.

// preemptive multitasking: OS scheduler can pause a running process
// even if it hasn't finished, to give CPU time to another process
if (higher_priority_task_ready) {
    save_context(current_task);
    switch_to(higher_priority_task);
}

Kubernetes: pod preemption — a higher-priority pod evicts a lower-priority one to get scheduled when resources are tight.

General English: “to preempt a question” = answer it before someone asks, i.e. act first to prevent something.

GPU/training context (relevant to you): preemptible/spot instances — cloud providers can preempt (kill) your training job on spot GPUs to reclaim capacity for higher-paying customers. This is why checkpointing matters when training on spot MI300X/A100 instances — if preempted, you resume from the last checkpoint instead of losing the run.

Back Donate