Spot Instances for Non-Critical Workloads
40% cost reduction on batch jobs — interruption handling, checkpointing, and what we never run on spot.
We moved report generation and log aggregation workers to spot node groups. Gateway stayed on on-demand.
Results (90 days)
| Metric | On-demand baseline | Spot |
|---|---|---|
| Monthly compute | $4,200 | $2,480 |
| Job interruption rate | 0% | ~4%/week |
| Failed jobs (no retry) | 0 | 12 (before checkpoint fix) |
Safe for Spot
- Idempotent batch processors
- Queue consumers with visibility timeout
- CI build agents (state in artifact store)
Never on Spot
- Gateway data plane
- Stateful databases
- Anything that cannot resume from last checkpoint within 2 minutes
cost , aws , spot · AWS , Kubernetes