Discussed generally in Why Cloud Bills Get Out of Control, on-demand pricing carries a real premium over committed pricing. Spot instances are a third option entirely — a steep discount, in exchange for accepting that the capacity isn’t guaranteed to stay yours.
What a spot instance actually is
Cloud providers maintain physical capacity sized for peak demand across all their customers, which means a meaningful amount of compute capacity sits unused at any given time, since actual demand rarely hits that peak everywhere at once. Spot instances — AWS’s EC2 Spot and Google Cloud’s equivalent Spot VMs are the best-known implementations — let customers use that unused capacity at a substantial discount compared to on-demand pricing, often a considerable fraction off the on-demand rate, in exchange for one specific condition: the provider can reclaim that capacity, with limited advance notice, whenever it’s needed for a customer paying full on-demand or committed pricing instead.
The trade-off, precisely
This isn’t a hidden or unpredictable risk — it’s the entire, explicit basis of the discount. A spot instance can be interrupted with only a short warning period, which is enough time for well-designed software to save state or shut down gracefully, but not enough time to treat a spot instance as though it has the durability guarantees of an on-demand or reserved instance. The discount exists precisely because the provider is selling something with weaker guarantees, not because it’s an artificially cheap version of the same product.
What actually makes a workload a good fit
Spot instances are a strong fit for workloads with a few specific characteristics:
- Stateless, or state that’s cheaply recoverable elsewhere. A worker process that can be safely killed and restarted without losing meaningful progress, or that regularly checkpoints its progress somewhere durable, tolerates interruption far better than one holding irreplaceable in-memory state.
- Flexible about timing. Batch processing, data analysis jobs, and rendering workloads that need to complete eventually, but not at a precise guaranteed moment, absorb an occasional interruption and retry without a user-facing consequence.
- Horizontally scalable across many instances. A workload already distributed, in the sense discussed in Serverless vs. Containers vs. VMs, across many interchangeable instances tolerates losing a few of them far better than a workload concentrated on one or two instances the entire job depends on.
- Genuinely cost-sensitive at scale. The discount matters most, in absolute terms, for large, sustained compute needs — training workloads, large batch pipelines — where a substantial percentage discount translates into a meaningful total saving.
What’s a poor fit
Anything requiring guaranteed availability at a specific moment — a production database’s primary node, a user-facing service with no tolerance for sudden capacity loss, anything without a real, tested plan for handling interruption gracefully — is a poor match for spot capacity, no matter how attractive the discount looks on a cost report. Using spot instances for workloads that can’t actually tolerate interruption doesn’t just risk cost savings; it risks the reliability targets discussed in Understanding SLIs, SLOs, and Error Budgets, for a category of workload where that risk was never actually necessary.
Designing for interruption, not just hoping to avoid it
Getting real value from spot capacity means designing for interruption as a normal, expected event, not an edge case: handling the interruption notice gracefully, checkpointing progress somewhere durable, and — for workloads that need to keep running continuously — often mixing spot instances with a baseline of on-demand or reserved capacity, so a wave of simultaneous spot interruptions doesn’t take the whole workload down at once. This is a genuine design decision made deliberately, similar in spirit to how chaos engineering treats failure as something to plan and test for directly, rather than an unlikely event to hope never happens.
Key takeaway
Spot instances offer a real, substantial discount in exchange for accepting that the provider can reclaim that capacity with limited notice — a trade that’s genuinely worthwhile for stateless, flexible, horizontally scalable, cost-sensitive workloads, and a poor fit for anything that needs guaranteed availability at a specific moment. The discount only pays off if the workload is actually designed to handle interruption gracefully, not just assumed to tolerate it.
This article explains general cloud compute pricing concepts; specific discounts, notice periods, and availability vary by provider and change over time. See our disclaimer.