Your Spot vs On-Demand Mix for Kubernetes, Explained

Spot can save 70–90%, but too much exposure causes interruptions and pager fatigue. Here is a practical way to set your spot/on-demand mix.

Classify workloads

• Spot-friendly: stateless web, async workers, batch jobs with retries.
• Mixed: internal services with moderate SLOs; run a blend of spot and on-demand nodes.
• On-demand only: latency-critical or stateful components (databases, control plane).

Set target mix per class

• Spot-friendly: 70–90% spot, rest on-demand buffer.
• Mixed: 30–50% spot with fast fallback.
• On-demand: 0–10% spot for non-critical replicas only.

Engineering controls

• Multiple spot instance types and AZs; diversify to avoid single-market volatility.
• Pod disruption budgets so interruptions don’t wipe all replicas.
• Taints/tolerations: keep critical workloads on on-demand; point spot-only workloads to spot pools.
• Node auto-repair that promotes on-demand capacity when spot evaporates.

Monitor the right signals

• Interruption rate per node group; keep it below 5% weekly for mixed workloads.
• Eviction impact: error rate or latency during interruption drills.
• Savings vs toil: track hours spent remediating spot churn; if toil climbs, reduce spot share.

Review quarterly

• Rebalance mix after big product launches or traffic shifts.
• Update allowed instance types; retire expensive or interruption-prone options.
• Recalculate per-region pricing to keep the savings real.

Spot is not “set and forget.” With clear classes, guardrails, and monitoring, you get the savings without turning SRE into an interruption helpdesk.***