Azure 2 min read 1

Spot Instances Are Cheap Until the Wrong Job Lands on One

A 70% discount with a two-minute eviction notice is a great deal for some workloads and a terrible one for others. How to tell which is which.

Spot Instances Are Cheap Until the Wrong Job Lands on One
Azure · October 5, 2026

Spot (preemptible) capacity is the best discount in cloud computing and the most common source of "why is my batch job stuck at 99% for six hours" tickets.

The deal is simple: up to 70–90% off, in exchange for the provider taking the capacity back with about two minutes' notice.

The question that decides everything

Not "can we afford it?" but "what happens when this node vanishes mid-work?"

Workload Spot-friendly? Why
Stateless web tier behind a LB Yes Eviction drains connections; replicas absorb it
Batch with checkpointing Yes Resumes from last checkpoint
Batch without checkpointing No Loses the whole run on eviction
Single-replica database Absolutely not Data loss and split-brain risk
CI runners Yes Jobs re-queue; idempotent by design
Long training run, no checkpoint No Days of compute, gone

The pattern that makes spot safe

Run redundant, interruptible capacity. The safety property you want is: no single eviction can lose work or drop a user request.

  • Keep a small on-demand floor for capacity that must exist.
  • Put the elastic burst on spot, spread across instance types and zones so one pool's reclamation does not take everything.
  • Make work checkpoint or re-queue. AWS eviction notices give you roughly two minutes — enough to drain, not enough to finish.
  • Handle the interruption signal explicitly: stop accepting work, flush, checkpoint, exit cleanly.

On Kubernetes this is exactly what node pools with mixed instance types plus a cluster autoscaler give you, and it is why "spot for everything" fails while "spot for the burst" saves real money.

Do the arithmetic honestly

If a job takes four hours and the pool reclaims every ninety minutes on average, an un-checkpointed job may never finish. The discount is only real if the work survives the reclamation.

Spot pricing is not a discount on compute. It is a discount in exchange for designing for interruption.

Teams that treat it as a pure cost cut end up paying for it in engineering time. Teams that treat it as a design constraint get the 70% and keep their weekends.

PB
Written by

Parul Bedi

Part of the Azure Gravy writing team.

Up next

Your Base Image Is a Supply Chain Decision, Not a Convenience

Docker October 10, 2026 1 min read
Open post

Conversation

1 comment
  • AN Ada Nwosu
    The workload table is the clearest summary of spot suitability I have read. Copied into our capacity planning doc (with credit).

Leave a comment

Replying to a comment.
Be kind. Markdown is not supported in comments.
Keep reading

Related posts

All posts