CloudCostCompareSpot and preemptible instances: cheap compute you can lose
Clouds keep spare capacity so they can absorb demand spikes. Spot instances let you use that spare capacity at a lower rate, with one condition: the vendor can take it back when it needs it, usually with only a short warning.
What each cloud calls it
| Cloud | Name |
|---|---|
| AWS | Spot Instances |
| Azure | Spot Virtual Machines |
| Google Cloud | Spot VMs (the older name was preemptible VMs) |
| Oracle Cloud | Preemptible instances |
The mechanics differ in the details: how much notice you get, whether the price moves with demand or is fixed, and how a stopped instance behaves. Read the vendor’s own page for the current rules, because they change.
Jobs that suit spot
Spot works when losing a machine costs you a little re-work, not a failure.
- Batch processing that splits into many small tasks, where a lost task is simply run again.
- Build and test farms.
- Rendering and encoding.
- Data processing and model training that save progress to storage as they go.
- Stateless workers behind a queue, where another worker picks up the message.
Jobs that do not
- A database with the only copy of the data.
- A customer-facing service with no spare capacity to take over.
- Anything that takes a long time and cannot save its state.
Designing for interruption
- Save progress to durable storage at regular points, so a restart resumes instead of starting over.
- Make tasks small and repeatable, so losing one is cheap.
- Listen for the warning. Each cloud signals an imminent stop; use it to save state and finish cleanly.
- Spread across sizes and zones. Capacity for one size in one zone can disappear while another is available.
- Keep a fallback. A small on-demand group can carry the minimum while spot adds the rest.
What it does to an estimate
Spot prices are not stable enough to put in a monthly estimate, and some vary by hour. Our estimates and the price catalog show on-demand prices only, so a spot plan will usually cost less than they say. Treat our figure as the ceiling for the work you move to spot.
Pairing with commitments
A common pattern is a committed floor for the steady part of the load (see reserved instances and savings plans), on-demand for what must not be interrupted, and spot for the flexible remainder. The overview is in cloud pricing models explained.