Spot instances offer the same compute at 60–90% lower cost. The catch? The cloud provider can reclaim them at any time. How do you deal with that?
How Spot Works¶
Cloud providers have unused capacity. Instead of letting it sit idle, they sell it at a discount. Azure Spot VMs: 30-second eviction warning. AWS Spot: 2-minute warning. Pricing fluctuates based on demand.
Suitable Workloads¶
- CI/CD build agents — build fails? Restart, no loss
- Batch processing — checkpointing, idempotent tasks
- Dev/test environments — tolerant of interruption
- ML training — with checkpoints and resume
- Stateless web workers — behind a load balancer, automatic replacement
Architecture for Spot¶
Kubernetes: dedicated node pool with a taint, workloads with tolerations. Mixed strategy: on-demand base (30%) + spot overlay (70%). Graceful shutdown handling — a SIGTERM handler saves state.
Our Savings¶
CI/CD on Spot: -73%. ML training: -68%. Dev environments: -80%. Overall impact on cloud bill: -22%.
Spot = Free Money (Almost)¶
If your workloads tolerate interruption, not using Spot is wasteful.
Need help with implementation?
Our experts can help with design, implementation, and operations. From architecture to production.
Contact us