Spot Capacity
also called Preemptible, Interruptible Instances
Using a provider's spare capacity at a steep discount, in exchange for the provider being able to reclaim it at short notice.
Discounts of 60% to 90% are available for accepting interruption, and the engineering question is simply which workloads can tolerate it.
Good candidates share a shape: work that can be checkpointed or retried without loss. Batch processing, CI runners, data pipeline stages, rendering, model training with checkpointing, stateless web tiers behind a load balancer where instance loss is already handled, and any queue consumer where an unacknowledged message returns to the queue.
Poor candidates: anything holding state that cannot be reconstructed, databases, long-running jobs with no checkpoint that would restart from zero, and anything where a two-minute warning is insufficient to hand off cleanly.
The practices that make it reliable in production. Diversify across instance types and zones, since interruption correlates within a pool and a fleet drawn from one type can be reclaimed together. Handle the interruption notice properly — drain, checkpoint, deregister — because the difference between a graceful and an abrupt loss is entirely in that handler. Mix spot with a baseline of on-demand or reserved capacity so that a large-scale reclamation degrades rather than removes service. And ensure the orchestrator reschedules rather than the workload being simply gone.
The economics worth stating to a finance stakeholder: spot is not a discount on the same product, it is a different service level. Buying it for a workload that cannot absorb interruption converts a saving into an incident.