Spot and Preemptible Instances: Cheap Compute With a Catch

What each platform calls it

Thirty seconds is enough to stop accepting work and checkpoint. It is not enough to finish a long job, which is why the workload has to be designed for it.

Workloads that fit

• Batch processing — image and video encoding, data transformation, report generation.
• CI and build runners — a lost build is re-run, not lost work.
• Machine learning training with checkpointing to storage.
• Stateless web tiers, mixed with on-demand capacity behind a load balancer.
• Analytics jobs on frameworks that already handle node loss.
• Development and test environments where an interruption is an inconvenience.

Workloads that do not

• Databases and anything holding state on local disk.
• Long jobs with no checkpointing, where an interruption loses hours.
• Anything with a hard latency commitment during peak.
• Single-instance services with no redundancy.
• Workloads on rare instance types with thin spare capacity, where interruption is frequent.

Designing for interruption

1. Handle the termination notice. Every platform exposes it through instance metadata. Poll it and react.
2. Drain gracefully — stop taking new work, finish or checkpoint what is in progress, deregister from the load balancer.
3. Checkpoint to durable storage, not to local disk, which disappears with the instance.
4. Make jobs idempotent, so re-running after an interruption is safe.
5. Diversify instance types. Spreading across several families and zones dramatically reduces simultaneous interruption.
6. Mix with on-demand. A base of on-demand capacity plus spot for the rest gives most of the saving with none of the availability risk.

That last pattern is the one most production estates use, and it is why spot is not only for batch work.

How likely is interruption

It varies by instance type, region and time. AWS publishes interruption frequency by instance type, which is worth checking before committing an architecture to a particular family.

General patterns: older and more common families interrupt less; scarce types such as high-end GPUs interrupt far more; and interruption clusters during regional demand peaks, so several instances of the same type may go at once. Diversification is the defence.

Spot versus commitment discounts

They combine well. Commit to the baseline, run steady growth on demand, and use spot for burst and batch.

Common questions

Where to go next

See reserved instances and commitments and cutting a cloud bill.