concept

Interruption Tolerance

The property that determines whether a workload can use heavily discounted pre-emptible capacity, defined by what happens when an instance is reclaimed at short notice.

Spot and pre-emptible instances offer discounts of 60–90% in exchange for the provider reclaiming them, typically with a warning measured in seconds to a couple of minutes.

A workload qualifies if losing an instance mid-execution is recoverable without data loss or a failed user request.

Good fits: batch processing, CI runners, data pipeline stages with checkpointing, media encoding, model training with checkpoints, and stateless web tiers behind sufficient on-demand baseline capacity.

Poor fits: databases and stateful services, anything with long non-checkpointed processing, and single-instance workloads where the interruption is the outage.

What makes it work in practice:

Handle the termination notice. Drain connections, checkpoint, requeue in-flight work. A workload that ignores the notice loses whatever it was doing.

Diversify across instance types and zones, since reclamation correlates within a pool. A request spread across six instance families is far less likely to lose everything at once.

Keep an on-demand or committed floor so the service survives a broad reclamation event.

The framing for a business conversation: spot converts a cost saving into an availability requirement — the architecture must already tolerate instance loss, which most well-built systems should anyway.